1. Introduction
Major cutting-edge and disruptive technologies are advancing rapidly worldwide. This advancement is driving a new wave of technological revolution and industrial transformation. With the rise of the digital economy, smart technology will bring about a revolution in technological paradigms and production methods. This will lead to a change in comparative advantages among nations and have a transformative impact on the international trade landscape. As a vital department dedicated to promoting global trade facilitation, Customs plays a pivotal role in accelerating customs clearance efficiency and stimulating world trade. The enhancement of ‘smart customs’ can contribute to the creation of a secure, convenient and efficient regulatory environment. Additionally, it can optimise regional supply chains, industrial chains and value chains, thereby improving the quality and growth potential of global trade. Therefore, many economies and organisations have explored the concept of smart customs and gained valuable innovative practical experiences worth sharing. The theme for the World Customs Organization (WCO) International Customs Day in 2022 was ‘Scaling up Customs Digital Transformation by Embracing a Data Culture and Building a Data Ecosystem’. This theme provided a blueprint for customs authorities to collaborate in developing an international customs ‘smart technology and service upgrade’. In response to changes in the global trade landscape and the promotion of trade facilitation, customs authorities across economies must strengthen the use of smart technology and innovative approaches. They should work together to create a smart customs system. Additionally, there is a need to improve information sharing and mutual recognition of regulation leading to the establishment of a smart border. Lastly, it is crucial to deepen cooperation with various stakeholders to collectively build smart connectivity.
The vision of the WCO is ‘Bringing Customs together for a safer and more prosperous world. Borders divide, Customs connects’ (WCO, n.d.). The WCO’s strategic paper Customs in the 21st Century, which proposes establishing a fair, impartial and equitable global economic governance system (WCO, 2008), defines the fundamental principles and key elements for future customs operations. The WCO is actively promoting ‘SMART borders for seamless trade, travel, and transport’ to guide the international customs community towards modernisation and to ensure the security and facilitation of global trade (WCO, 2018). In response to the call of the WCO, China Customs has taken inspiration from successful practices of other customs authorities worldwide and has made continuous efforts to advance its modernisation, considering its own unique circumstances (General Administration of Customs Control People’s Republic of China [GACC], 2020). The objective is to adapt to the latest technological changes and new modes of international trade and to promote trade facilitation, ensure supply chain security and uphold the international free trade system and open global economy (GACC, 2020). To achieve these goals, China Customs has developed and released the initiative ‘Smart Customs, Smart Borders, and Smart Connectivity’ (GACC, 2020). The ‘Three Smart’ customs cooperation system is a comprehensive and interconnected project with smart customs serving as its foundation, smart borders as its expansion and smart connectivity as its ultimate vision (Q. Li, 2022).
Smart customs refers to the utilisation of information technology (IT), network technology, big data, cloud computing, AI and other emerging technologies. It involves the use of intelligent hardware, automation equipment and other innovative tools to enhance the capabilities of information collection, risk assessment, precise deployment and control, as well as non-intrusive inspection (Cao et al., 2022). This enables digital processing, networked transmission, automated operation and intelligent decision-making in customs supervision and control processes, leading to improved efficiency. The aim is to accurately prevent and combat illegal activities while ensuring smooth customs clearance for legitimate goods. The construction of smart customs is a complex and time-consuming task, requiring continuous improvement, adaptation to external changes and innovation in IT (J. Q. Wang et al., 2020). The path to smart customs is based on an in-depth exploration of how to build a smart customs control and service system.
After almost four decades of IT development, China Customs has made significant progress in the implementation of smart customs. The automation of the entire business process is achieved through the smart customs operating system, as depicted in Figure 1. This system enables 100 per cent electronic declaration. However, approximately one per cent of the declarations are returned due to non-compliance with the filing specifications. Among these, 75 per cent can be processed automatically, while the remaining 25 per cent require manual review. This is because the documents accompanying the customs declaration, such as the certificate of origin, contain unstructured data. Customs declarations that have successfully passed the examination undergo a risk screening process. Due to limited resources for on-site inspection, approximately four per cent of the declarations are selected for physical inspection. To ensure fairness, one per cent of these inspections are chosen randomly for double inspection. However, due to a shortage of reviewing experts, only about 50 per cent of the actual cargo inspections are conducted manually, with the remaining 50 per cent inspected using machines. Statistically, around 20 per cent of the cargo inspected during live cargo checks exhibits abnormalities (with a seizure rate of 20 per cent) and requires special handling. Normal cargo is released for subsequent management.
The Customs and Excise Department of China Customs collaborates with various departments such as State Taxation, Market Supervision, Commerce, Health, Agriculture, Banking and Foreign Exchange through network verification to conduct comprehensive management. However, there are three main challenges in the development of smart customs. Firstly, the level of electronic audit intelligence is limited, resulting in a high proportion of manual document checking. This not only consumes significant human resources but also hampers customs clearance efficiency (Zeng, 2021). Secondly, limited by the accuracy of risk screening and deployment control, the on-site seizure rate is low. This leads to wastage of customs inspection resources, increased trade compliance costs and reduced deterrent effects on smuggling (Zhang et al., 2020). Lastly, customs machine inspection and judgement lack sufficient human expertise, resulting in a low proportion of machine inspection, and loopholes in intercepting prohibited and restricted goods (X. C. Wang, 2024). These loopholes pose a threat to national security at the country’s gates. Therefore, it is crucial to overcome these three bottlenecks to address the main challenges in customs operations.
In recent years, China Customs has made efforts to overcome these challenges by actively exploring the integration of AI and other technologies with front-line supervision. One successful development is Intelligent Customs Inspection (ICI) in China, an application that uses AI to review scanned images of imported and exported containers and baggage (Z. Y. Huang et al., 2024). This machine-assisted tool can even replace human intervention in some cases. The intelligent inspection application effectively addresses the main challenge faced by Customs, which is to balance regulation and facilitation. It plays a crucial role in combating smuggling, ensuring national border security, improving the business environment and enhancing customs supervision capacity.
This paper focuses on the application of intelligent inspection developed by China Customs. Primarily, it discusses the technical principles of AI used in its application and analyses the deployment of ICI, along with its impact on efficiency, based on relevant data provided by China Customs. The paper also discusses the prospect of ICI including a paradigm study of security screening technology based on X-ray multi-characteristic imaging, the design concept of the customs intelligent inspection information platform and the potential applications of intelligent customs document inspection. The goal of this paper is to contribute to the construction of global smart customs by providing valuable insights and exchangeable Chinese programs.
2. Literature review
2.1. Artificial intelligence technologies
The concept of ‘machine intelligence’ has gained popularity due to the Turing test (Vidyasagar, 1990). It refers to a program designed to simulate a child’s learning process and gradually develops into a thinking program with intelligence comparable to that of an adult brain. Through continuous development, AI research has evolved into three main areas (Cortes & Vapnik, 1995): logic-based knowledge graphs (Lee & Wong, 2017), deep learning based on a controller (Chen & Chan, 2021) and deep learning based on neural networks (Huber & Imhof, 2023). In 1998, a computer vision algorithm called LeNet, which utilised the convolutional neural network, was proposed for character recognition (Lecun et al., 1998). Computer vision tasks entered the era of convolutional neural networks in 2012 when researchers utilised AlexNet, a model with seven hidden layers, to emerge victorious in the ImageNet large-scale image classification competition (Krizhevsky et al., 2017). In 2014, scholars from the Visual Geometry Group at the University of Oxford presented a paper introducing a series of 19-layer-deep ‘VGG’ models (Simonyan & Zisserman, 2014). The paper proposed extending the receptive field (RF) by utilising a small convolutional kernel combined with stacked convolutional layers. This approach not only reduced the number of parameters in the deep convolutional network but also introduced new nonlinearities to the model (Simonyan & Zisserman, 2014). Subsequently, other scholars introduced the ResNet family of models (Girshick et al., 2014). These models enable the training of deeper convolutional neural networks by reducing the size of feature maps, thereby transforming the training targets into residuals. In response to the challenge of dealing with numerous candidates when transitioning to deep neural networks for feature vector extraction, scholars have proposed the Region-based Convolutional Neural Network (R-CNN) algorithm (Girshick et al., 2014). Subsequently, an improved version called the Fast R-CNN algorithm has been proposed (Girshick, 2015). To address the computationally intensive two-stage target detection algorithm, Liu et al. (2016) proposed a single-stage detection algorithm called Single Shot Detector (SSD) with You Only Look Once (YOLO). Currently, CNN-based single-stage algorithms remain the dominant technology for image target detection.
Even though current state-of-the-art AI systems have achieved comparable or even superior performance to humans in certain domains, they still face significant limitations in their lack of comprehension, which is rooted in the ability to understand meaning. This comprehension barrier poses a challenge for AI and necessitates the exploration of approaches that can enhance the field by equipping AI with common sense through the utilisation of knowledge, abstraction and analogies (Koivisto & Grassini, 2023). Throughout the vicissitudes of over 60 years, the term ‘artificial intelligence’ has been surrounded by numerous controversies, yet a universally accepted and agreed-upon definition remains elusive. Despite decades of exploration, there is still no precise definition that has gained unanimous acceptance. It is unlikely that a precise definition of AI will ever be agreed upon by all (Grosz et al., 2016). A comprehensive definition of AI is crucial to establish a shared understanding of the concept within the public sector and will serve as a fundamental basis for conceptualising the various applications and challenges associated with AI (Wirtz et al., 2018). A comprehensive definition of intelligence encompasses the capacity to effectively communicate with others, acquire knowledge, assimilate and apply empirical information, and navigate through uncertain situations (Legg & Hutter, 2007). Simultaneously, there are studies that interpret the term ‘artefacts’ as replicas generated by humans (Nandhakumar & Aggarwal, 1985). In recent years, scholars such as Simon (1977) have increasingly focused on studying and exploring the concept of ‘AI’ from a comprehensive standpoint. They view it as a research field that aims to develop intelligent machines and software capable of replacing human thinking and decision-making (Simon, 1977). Advances in the field of AI are resulting in a new level of computation. These advances enable systems to act as autonomous agents and learn independently. Additionally, they can assess their environment and think in terms of values, motivations and emotions (Barth & Arnold, 1999). It has been argued that the study of AI focuses on enabling machines to learn various aspects of simulation and accurately characterise different features of intelligence. This can be described as the science and engineering of making intelligent machines, particularly intelligent computer programs (McCarthy et al., 2006). According to scholars such as Wirtz et al. (2018), AI is defined as the capability of a computer system to exhibit human-like intelligent behaviours. These behaviours are characterised by various core competencies such as perception, understanding, action and learning. In terms of AI implementation, it is worth noting that AI, machine learning (ML) and deep learning are often used interchangeably. However, it is crucial to understand that they are distinct concepts, although they are interconnected. Deep learning is a subset of ML, which in turn is a subset of AI (Mikhaylov et al., 2018).
The impact of AI on human society is becoming increasingly evident. It is now impossible to ignore its impact on management by governments, which has garnered significant attention from scholars. The relationship between non-programmed decision-making and AI has been explored by Simon (1996). Simon argues that the problem-solving mechanisms of the human brain are structured similarly to the symbolic logic systems of computers. This is because both computers and humans learn through trial and error (Simon, 1996). Simon (1983) further points to AI as one of the technological tools for decision-making and public policy and emphasises the importance of IT for public utilities and public policy (Simon et al., 1991). Galloway and Swiatek (2018) figured that current discussions have focused excessively on AI’s capacity for automating tasks, overlooking its wider technological, economic, and societal impacts on public relations. In recent years, scholars (Wirtz et al., 2018) have highlighted the significance of considering the impact of AI in the public sector. Despite the growing investment in AI research and the increasing number of research contributions, it is important to recognise that AI for public use is still a relatively new area of study, more efforts are needed to thoroughly describe the relevant applications and challenges in this field. Several scholars, such as Kohli et al. (2019), have highlighted that the United States (US) has started acknowledging the significant value of AI in the public domain. As a result, the country has initiated numerous AI initiatives with substantial investments, uncovering a diverse range of potential application areas. In 2018, the United Kingdom (UK) Government released a Sector Deal between government and the AI sector (UK Government, 2018). After adhering to a liberal approach for an extended period, the US Government took the initiative to promote AI development and it requested national agencies, excluding the defence sector, to invest in AI to meet the growing public demand (The White House, 2023).
The previous research highlights the interdependence of AI development and application on the contributions of technology developers, social organisations, and other stakeholders. Notably, the government should also play a crucial role in this process.
2.2. Customs clearance supervision
To conduct intelligent customs research effectively, it is important to begin with a thorough analysis of key business areas such as customs clearance operations, cross-departmental networking, digitalisation of business decision-making, informatisation of governmental affairs system, and informatisation of logistics monitoring and control. This analysis should aim to identify and anticipate a range of benefits and effects brought about by customs informatisation construction, including quality enhancement, efficiency improvement and strengthened supervision (Yang et al., 2009). The role of Information and Communication Technology (ICT) is crucial in the operations of modern customs administrations. The use of computerised systems has greatly expedited the processing of information, leading to a more efficient and streamlined approach. Furthermore, ICT has empowered customs administrations to employ a risk management strategy in the selection of goods and passengers, thereby enhancing security measures (Choi, 2011). Additionally, ICT has facilitated improved compliance by the private sector and enhanced service delivery to businesses (Choi, 2011). Various customs administrations have implemented different approaches to the application of big data. For instance, US Customs and Border Protection (US CBP) has introduced a centralised location to apply big data technologies to its own data (US CBP, 2019). Similarly, New Zealand Customs Service and its border partner, the Ministry of Primary Industries, are currently engaged in a modernisation project that involves the use of analytics tools for risk assessment and border management (Okazaki, 2017). The WCO has begun researching potential case studies and applications of blockchain technology for customs and other border agencies (WCO and World Trade Organization [WTO], 2022). The objective is to enhance compliance, trade facilitation, and fraud detection, including the prevention of illicit trade through the misuse of blockchains and bitcoins. Necessary changes in legal and regulatory frameworks are also considered (Okazaki, 2018). Governance by data is a growing global trend, which is supported by strong national public policies. The foundation of these policies lies in open data, AI, and decision-making supported by algorithms. However, despite this trend and some technical advances, Customs still faces obstacles in implementing ambitious data use policies (Mikuriya & Cantens, 2020). As evident from previous scholarly discussions, the enhancement of customs supervision through intelligent means is a comprehensive and evolving endeavour that necessitates the seamless integration of advanced IT.
In recent years, customs agencies worldwide have been actively exploring ways to promote digital transformation and intelligent upgrading. In 2020 SMART Customs Initiative, Japan Customs aims to advance the digitisation of customs functions/activities using various state-of-the-art technologies. The goal is to foster the growth of trade, ensure a safe and secure society, and create a prosperous future (Customs and Tariff Bureau, Ministry of Finance of Japan, 2020). The Business Transformation and Innovation Division of the US CBP is responsible for evaluating and analysing the application scenarios of new technologies in the CBP’s business needs (US CBP, 2019). They have implemented cutting-edge intelligent technologies such as AI, automatic categorisation technology, and distributed e-commerce product information storage technology to address the challenges of inaccurate data and high customs supervision costs associated with cross-border e-commerce products (Canada Border Services Agency [CBSA], 2018). The Thai Customs Department showcased their research and development in customs supervision at the WCO Asia-Pacific Regional Seminar 2021 whereby it explored the application of X-ray technology, radio frequency identification (RFID) technology, electronic locks, and other advanced technologies to enhance customs informatisation (Asia Pacific Economic Cooperation Sub-committee on Customs Procedures [APEC SCCP], 2024). The Korea Customs Service (KCS) is currently investigating the development of an advanced passenger risk analysis system that utilises big data analysis technology (KCS, 2019b) and is considering implementing a blockchain-based mechanism (KCS, 2019c) to exchange data on electronic transactions between countries, as well as exploring the use of an AI-based customs X-ray inspection system (KCS, 2019a). CBSA has investigated the concept of a ‘secure corridor’, which is a technological corridor that incorporates various sensors and remote processing of information related to transportation (CBSA, 2018). These include RFID sensors, multiple camera views (such as driver, licence plate and trailer), voice-to-voice communications, risk analysis and other technologies. The objective of this corridor is to reduce clearance times for ‘trusted/pre-approved’ carriers (CBSA, 2018).
2.3. Customs intelligent inspection based on AI
There is significant potential for the global development of AI. Governments worldwide are recognising the strategic importance of AI. While current customs applications rely on many X-ray machines for supervision, the growing need for stricter security measures has led to a demand for new computed tomography (CT) imaging equipment (WCO and WTO, 2022). By examining the current state of contraband identification in the security industry, Mouton and Breckon (2015) highlight the research direction and potential of customs in CT image target detection. In addressing the issue of limited identifiable target types in image representation and foreign object detection, a framework proposed by Andrews et al. (2017) utilises supervised convolutional neural networks to surpass the constraints of current technology, which struggles to identify unfamiliar contraband types. The results showcase the efficacy of the framework on a scanner image training dataset. Andrews et al. (2017) discusses how other scholars have systematically summarised the algorithms utilised in image preprocessing and image understanding, with a focus on intelligent recognition in different cargo scanning devices. The objective is to improve customs inspection procedures through the integration of smart identification technology (Rogers et al., 2017). Additionally, the current limitations of the algorithms are assessed and potential future development paths are explored. Researchers have rapidly established connections between customs issues and highly advanced areas of AI, such as image recognition applied to non-intrusive inspections (NII) (Jaccard et al., 2017; Kolokytha et al., 2017).
To investigate the obstacles faced in the progression of intelligent machine inspection, Ye et al. (2018) utilise AI technology to pinpoint deficiencies in traditional management approaches and make recommendations for organisational restructuring and process innovation. The European border agency, Frontex, employs a geographic information system to effectively oversee and manage the external borders of the European Union (EU). This system integrates data from various sensors and sources including ships, individuals and other databases (Malinowski, 2019). Q. Wang et al. (2020) discovered that utilising various mainstream network architectures, such as ResNet in different scales, along with implementing diverse data augmentations like occlusion of objects, random flipping and rotation, can result in an average accuracy (mean average precision, mAP) of 65.3 per cent in prohibited object recognition.
As mentioned above, in the context of customs clearance supervision, most of the research literature focuses on the intelligent utilisation of customs clearance machine inspection equipment and the exploration of imaging technology. To investigate methods for enhancing customs clearance efficiency and supply chain security, Tuszynski et al. (2013) proposed utilising an algorithm to compare scanned images of container goods with non-suspicious historical images of the same category. This process can determine if the goods in the container match the declared types. The experimental results applying this approach to the US CBP showed promising effectiveness, with confidence levels aligning with declared types, indicating potential for aiding customs officials in identifying unusual shipments. To achieve automatic identification of key categories such as cigarettes, weapons and drugs in containers, and facilitate automated data sharing among customs authorities, Kolokytha et al. (2018) proposed an integrated system based on intelligent identification. This system aims to integrate intelligent identification algorithms into the inspection process, enabling efficient identification and detection. Malarvizhi et al. (2021) proposed a system that integrates manifest data of goods and transportation vehicles with scanned X-ray images to achieve intelligent matching. If the goods do not match, the system will prompt customs officers to detain the goods for further inspection. Several studies, such as Memmel et al. (2021), have suggested a slice re-fusion method to address the challenges in security CT contraband recognition. This method primarily focuses on resolving the issues related to the extensive computational resources required for processing 3D data and the scarcity of labelled training data.
As evidenced above, customs authorities worldwide are actively pursuing digital transformation and intelligent upgrading (APEC SCCP, 2024). They are prioritising technological innovation and its application in business areas, tailored to their unique characteristics. Developed countries are rapidly promoting the application of technology, leading to fierce competition. China Customs should seize the opportunity presented by emerging disruptive technology development and contribute China’s strength to the efficient implementation of the ‘Three Smarts’ initiative. The 2022 report on disruptive technologies jointly published by the WTO and the WCO (WCO and WTO, 2022) demonstrates the willingness and practice of Customs to adopt disruptive technologies. This report aims to identify how these advanced technologies can enhance trade facilitation and help customs agencies achieve their objectives of ensuring national security at borders and fair revenue collection.
3. Innovative practices in the supervision of customs clearance inspection
3.1. Risk system of customs clearance
The port environment has become increasingly complex, resulting in challenges for customs clearance transactions (Q. Li, 2022). The customs business supervision process is characterised by dynamic changes, high volume and complexity. In addition to structured data, there is also a significant amount of unstructured data, including text, video, images and voice. This requires the utilisation of various intelligent services to effectively handle and integrate these different data types into the supervision process (Z. Y. Huang et al., 2024). To tackle the complexity of risk screening in customs declarations, AI and big data technologies play a crucial role. These technologies can be utilised to develop algorithms that enable efficient and accurate intelligent review of images and accompanying documents, as illustrated in Figure 2.
3.2. ICI in China
3.2.1. Definition of ICI and machine inspection
Combining AI with customs supervision transactions, this technology utilises ML and expert knowledge to analyse historical images and information of goods and articles. It then forms automatic identification algorithms for machine inspection images of H986 (an X-ray-based container scanning machine), CT and other equipment (Zhang et al., 2020). By combining the characteristics and attributes of goods, articles and means of transport, this technology automatically screens corresponding machine inspection scanning images. It assists in the manual judgement of these images and, through continuous optimisation, aims to replace human involvement with machines in the field of machine inspection. This approach is referred to as ICI in China (J. Q. Wang et al., 2020).
H986, CT machines and X-ray machines differ in their imaging principles, resulting in significant variations in the image data they produce. H986 and X-ray machines generate two-dimensional images, either single- or double-view, while CT machines produce three-dimensional images with a 360-degree view (Ye et al., 2018). This disparity in imaging capabilities leads to distinct amounts of data and information for different types of inspection equipment. CT and X-ray machines find wide application in travel inspection, express delivery and mail regulatory venues for overseeing packages and small cargo (X. B. Li et al., 2022). On the other hand, H986 is extensively used for supervising and inspecting bulk cargoes in land and sea transportation. Thus, while CT and X-ray machines are commonly employed in travel inspection, express delivery and mail regulatory venues, H986 finds its primary use in the supervision and inspection of bulk cargoes in land and sea transportation (Zhao, 2012).
3.2.2. Key technologies for ICI
ICI in China is primarily conducted using deep learning technology, such as H986. The container image is divided into three parts: the body, box, and loaded cargo (Z. Y. Huang et al., 2024). The body part and box part are analysed using silhouette algorithms to check for any abnormalities like folders or layers. The loaded cargo is examined at the level of individual goods, using the commodity code and its Chinese and English name as reference. A large quantity of historical images of correct commodities are used to extract various feature values and create models. These models are then used to verify if the declared goods match the actual goods mentioned in the electronic declaration. The models also help identify any attempts to deceive or significantly conceal the goods (Z. Y. Huang et al., 2024).
The concept of neural networks first appeared in 1943. In the computer field, deep learning is often used in algorithm improvement. Strictly speaking, the emergence of deep learning is due to the development of neural networks (Tian & Wang, 2019). In 2006, Hinton and colleagues published a seminal article on neural networks introducing the deep belief network, a successful multilayer training method (Hinton et al., 2006). This method utilises a hierarchical initialisation approach to effectively train deep neural networks, addressing the training difficulties encountered with the Back Propagation algorithm. As a result, neural networks have completed the transformation from shallow networks to multilayer deep networks, and the concept of deep learning was born, and it is increasingly used in various fields. This method sparked a research boom in neural networks, rapidly establishing itself as a key technology in the field of AI. Deep learning works by constructing a network model with numerous layers containing bionic neurons (Topol, 2019). The network is then defined with input data, output data and an error function. During the training process, the error between the output results and the labelled information is calculated repeatedly, allowing for autonomous optimisation of the network model’s parameters. The learned network structure and parameters are then used to capture the complex relationship between the input and output (Topol, 2019).
Figure 3 illustrates that the training process of a deep learning model follows a closed loop structure. This loop encompasses various stages such as data acquisition, algorithm development, optimisation and iterative refinement, as illustrated in Figure 4. The aim is to continuously explore and exploit the capabilities of the equipment, enhance algorithm performance and expand the range of applicable business scenarios (Cao et al., 2022). By doing so, the ICI system can effectively implement and enhance algorithms based on business requirements and data-driven insights.
The use of deep learning technology in ICI applications involves the core algorithm, which utilises hinge loss and other training functions. The algorithm primarily employs support vector machine (SVM) in the objective function to address classification problems such as ‘maximum interval’. The central idea is to analyse the input data and perform classification operations (Cortes & Vapnik, 1995).
The generic binary categorisation expression can be represented as where y is the predicted value that ranges between −1 and 1, and t is the target value that takes values of 1 or −1. This equation implies that achieving the correct classification of a given sample is sufficient, and it is discouraged to have excessive confidence, as there will be no reward when the sample’s distance from the dividing line exceeds 1. Therefore, the classifier discourages overconfidence and focuses on minimising the classification error (Cortes & Vapnik, 1995). If the classification is error-free, the loss is 0, otherwise the loss is 1-t y. Simulation of the SVM linear binary classification idea was conducted using MATLAB Release 2016b simulation software (MathWorks, 2016). As a general-purpose numerical simulation software, MATLAB has powerful computational and simulation capabilities (Gong et al., 2022). The results of this simulation are shown in Figure 5.
The ICI application can accurately identify contraband in containers or baggage parcels by utilising a combination of the complex variant of the classification detection algorithm, image semantic segmentation and other algorithms (Z. Y. Huang et al., 2024)
3.2.3. Deployment analysis of intelligent customs machine inspection
China Customs has progressively expanded the implementation of ICI, achieving comprehensive coverage of the system across various types of machine inspection equipment and inspection operation scenarios.
In the context of H986, China Customs currently utilises this equipment for a wide range of freight business scenarios, including maritime, highway and railroad operations (Ye et al., 2018). By analysing the distribution of H986 ICI deployment, which covers 31 customs administrations, the following characteristics are revealed (Figure 6). First, the deployment across nationwide maritime port business shows a relatively balanced coverage across all 31 customs administrations. However, there are differences in the deployment of H986 in highway port business among these administrations, and the deployment of H986 in railroad port business is relatively limited in scope. Second, the southern region has a higher concentration of maritime port business deployments, while the northwestern part of the country shows a greater concentration of highway port business and railroad port business deployments. Therefore, to enhance the coverage of railroad port business, it is recommended to expand the future deployment of intelligent image review application using H986 equipment. Additionally, a coordinated deployment of various business scenarios should be considered to achieve a balanced application in each customs area.
In terms of CT, China Customs utilises a comprehensive range of CT equipment that caters to various business scenarios, including express, mail, cross-border e-commerce and travel inspection (Ye et al., 2018). By analysing the distribution of CT equipment deployment, it becomes evident that the current deployment profile of CT equipment ICI covers 40 customs branches/offices in the following manner (Figure 7). First, the nationwide deployment status of express mail business shows a more balanced coverage among the 40 customs branches/offices. The deployment status of travel inspection business and mail business is relatively similar, while the deployment range of cross-border e-commerce business is smaller. Second, the northwest region has a higher deployment of travel inspection business, while the southern region has a higher deployment of express mail business, and the southern region also has more deployment of cross-border e-commerce business. Consequently, it is recommended that the future deployment of intelligent image review application of CT equipment be enhanced to increase the coverage of cross-border e-commerce business. Additionally, each customs area should coordinate the deployment of various business scenarios to achieve a balanced application.
3.2.4. Efficiency analysis of intelligent customs machine inspection
Since the full implementation of the ICI system, the Customs Department has experienced a significant increase in the number of seizures and an improvement in its supervision capabilities. Using H986, a total of 3,000 consignments with a value of approximately CNY6 billion were seized, while with CT, 1,700 consignments with a value of approximately CNY100 million were seized.
Figure 8 illustrates that there has been a notable progression in the development and implementation of ICI technology. This advancement has resulted in a substantial improvement in recognising various types of declared commodities and a significant increase in effectively intercepting these commodities.
Figure 9 indicates that the current percentage of CT devices is higher than the percentage of H986 devices in both recognised declared goods and effectively intercepted goods.
The current average recognition time for ICI of H986 equipment is approximately 10 seconds, while the average recognition time for ICI of CT equipment is approximately 5 seconds. This demonstrates efficient and accurate recognition.
4. Prospects for ICI
4.1. ICI informatisation platform
To facilitate extensive adoption and implementation, technical specifications have been developed by the authors for eight types of supervisory equipment, including H986, CT, X-ray machines and millimetre wave devices. These specifications aim to achieve unified equipment access, image format, data interface standardisation and system security. In line with this, China Customs has initiated the development of an informatisation platform for ICI application. The platform underwent a pilot test in June 2021 and is now operational. Its main purpose is to provide technical support for ICI in areas such as image management, algorithm management and data resource sharing.
The platform aims to achieve personalised algorithm package customisation for different customs districts based on business requirements, as shown in Figure 10. It also facilitates algorithm package issuance and maintenance.
To achieve efficient customs clearance while ensuring supervision, it is important to balance the relationship between ‘supervision’ and ‘facilitation’. This can be done by optimising and improving the efficiency of intelligent supervision provided by machine inspection equipment (Cao et al., 2022). Enhancing the efficiency and precision of automatic linkage comparison of charts and documents can reduce the workload of customs officers and shift from manual labour to intelligent systems (Shenzhen Customs Project Group, 2021). Therefore, it is necessary to explore innovative applications of ICI in the field of document identification.
4.2. Intelligent customs document inspection
4.2.1. Status of document recognition
Among the accompanying documents submitted by the import declarant, there are still 11 types of documents that have not yet been fully integrated into electronic networking. Therefore, it remains necessary to upload scanned copies of the relevant certificates of origin. In 2020, the total amount of taxes involved in these documents amounted to approximately USD1.6 trillion. Taking the certificate of origin as an example, China implemented 18 preferential trade arrangements in 2020, granting tax concessions totalling approximately CNY83.2 billion based on various types of certificates of origin submitted by importers. In 2021, there were a total of 19 certificates of origin under preferential trade arrangements, however, only nine of them have achieved electronic networking, while the remaining half are yet to be fully integrated into electronic networking.
4.2.2. Prospective analysis of intelligent customs document inspection
The customs computer system for unstructured documents with intelligent discrimination ability needs improvement. Additionally, there has been a significant reduction in incidental audit or manual intervention at the customs clearance site (Shenzhen Customs Project Group, 2021). However, different countries and enterprises have accompanying documents in different formats and texts, which may include official seals and multiple signature records. Without a customs clearance system to verify the validity of these scanned documents, relying solely on manual audit cannot fully meet transaction needs (Shenzhen Customs Project Group, 2021). This creates a risk of importers uploading incorrect or false evidence to fraudulently benefit from the system.
The total number of customs declarations from January 2020 to January 2021 was approximately 7,348,298,985, while the total number of non-manual interventions accompanying customs declarations during the same period was approximately 9,989,736,565.
The percentage of customs declarations without manual intervention was approximately 76 per cent of the total number of declarations in the year analysed, while the percentage of unstructured accompanying documents without manual intervention was nearly 89 per cent of the total number of accompanying documents. The high proportion of non-manual interventions may generate numerous unknown risks. Therefore, it is important to explore the integration of ICI in the field of document identification for unstructured documents accompanying customs declarations, which have a significant proportion of non-manual intervention. This integration can enable the implementation of intelligent document review, enhancing the ability to prevent unknown risks in customs declaration documents, reducing the rate of misjudgement in customs clearance risk screening, and achieving smart, efficient and accurate customs supervision. These advancements will contribute to the development of smart customs and bring about significant reform, innovation and practicality.
4.3. ICI based on X-ray multi-characteristic imaging
According to hedonic price theory, some scholars, such as Ji et al. (2021), proposed a multi-characteristic product analysis concept that can be extended to multiple fields, to provide suggestion and guidance for the decision-making and behaviour of relevant stakeholders. The current X-ray machine inspection equipment relies on the absorption of X-rays by materials for imaging, however, it faces challenges when it comes to detecting weakly absorbing substances such as human genetic material and drug powders, which consist mainly of carbon, hydrogen, oxygen and nitrogen (Mouton & Breckon, 2015). The traditional imaging technology has reached its limitations and cannot effectively detect these substances. X-ray multi-characteristic imaging technology exploits the wave-particle duality of X-rays. By introducing micro-nano grating to modulate the optical field, it can simultaneously capture three complementary physical properties of substances: absorption, phase and small-angle scattering. This approach significantly enhances the amount of information obtained about the substance (Z.-F. Huang et al., 2009). The phase information enhances the image contrast of weakly absorbing substances, while the small-angle scattering information is highly sensitive to the micro- and nano-structures of substances. This sensitivity allows for better recognition of different textures, such as liquid and powdered contraband, and compensates for the limitations of traditional absorption imaging (Partridge et al., 2022). The application of X-ray multi-characterisation technology in the field of security screening is rapidly growing and has shown significant potential (Looker et al., 2020). This technology has also improved the accuracy of image recognition and the effectiveness of material recognition, providing a solid foundation for the development of ICI.
5. Conclusion
This paper has presented an analysis of the current customs clearance operations of China Customs. In particular, it examined how China Customs has applied AI technology to the supervision transaction of customs machine inspection equipment, leading to the development of the ICI system. Additionally, the current deployment of ICI in China’s customs clearance machine inspection equipment and the resulting improvement in regulatory efficiency was discussed. Furthermore, the potential innovative applications of ICI in the future were explored, specifically in the areas of informatisation platform construction, intelligent document review and X-ray multi-characteristic imaging. The aim is to provide solutions from China for the future development of smart customs.
Acknowledgments
This research is supported by the Shanghai Science and Technology Project (No. 22YF1415400). Additionally, it is funded by the Young Teachers’ Research Initiation Project of Shanghai Customs College (No. 2315015A2020).