Since its inception, artificial intelligence has undergone multiple technological changes and large-scale commercialization, from early expert systems and machine learning to the ongoing popularity of deep learning and large models. With the rapid progress and continuous maturity of computing power, algorithms, and software platforms, industry has gradually become a key exploration direction of deep learning technology, and the application of industrial AI intelligent technology has emerged.

The industrial defect detection method based on deep learning can reduce the cost of traditional manual quality inspection, improve the accuracy and efficiency of detection, and thus play an important role in intelligent manufacturing.

Traditional machine learning and deep learning techniques rely on a large amount of annotated data and train models with excellent performance and certain generalization ability under supervision. However, as the perception environment and application scenarios change, the training of the model may encounter the following issues:

(1) Lack of defect samples, limited quantity and types of defect data in the production process, imbalanced data, and inability to exhaustively list the types and forms of defects in the production process.

(2) The annotation and cleaning cycle of data is long, requiring a lot of manpower and material resources.

(3) The performance of a well trained model will significantly decrease, and the cost of retraining the cycle is high.

(4) Cross domain learning and feature level data fusion of multimodal data.

The above issues have become obstacles to the implementation of industrial AI. How to solve problems such as data annotation efficiency, cross domain learning, and data management, and train models with more generalization, robustness, and scene adaptability have become a common issue faced by the academic and industrial communities.

For some scenarios that cannot be effectively solved by traditional methods, such as detecting small defects and defects, sorting objects in unstructured environments, etc., it can be classified as a "low factor high complexity" problem, which is an important field where deep learning plays an important role. Currently, it is also a widely used scenario in industrial AI. With the increase in computational complexity of scene mechanisms, deep learning can play a greater role.

In order to improve the implementation efficiency of deep learning in the industrial field and reduce the project implementation and deployment cycle, Huahan Weiye mainly explores and practices technology from the following aspects:

(1) Defect data generation: Using artificial intelligence technology to automatically generate defect simulation data, establishing a mapping relationship from the real world to the digital world based on AIGC technology, efficiently and perceptively digitizing the physical properties of workpieces in the real world (such as object size, texture, color, etc.), and generating multiple attribute samples based on a small number of samples, thus solving the problem of defect sample scarcity.

(2) Data management: In the production process, data from multiple production lines and workstations needs to be managed, requiring manual control of data. There is a lack of a data management system, which facilitates subsequent inheritance and continuous training. Huahan Weiye utilizes digital technology to achieve systematic data management and control across multiple workstations and scenarios, reducing the impact of human factors on data management and control.

(3) Data annotation: Currently, supervised learning is still the technical direction for industrial AI implementation. In order to improve annotation efficiency and reduce annotation time consumption, Huahan Weiye starts with interactive annotation to improve annotation efficiency,

(4) Multimodal data fusion: In industrial production, many defects cannot be captured from a specific angle or a single sensor. It requires multi angle light source illumination and multi sensor collaborative shooting to achieve visualization of multiple defects. In order to improve the fusion of multi angle and multi pose image feature levels and improve the accuracy of defect detection, Huahan Weiye has explored various technologies such as multimodal feature fusion and feature fusion based on image data streams to enhance the generalization performance of the model.

(5) Reduce sample data dependency: In order to reduce the dependence on samples during the training process and improve the model‘s adaptability to different production lines and scenarios, Huahan Weiye learns from small samples, transfer learning, and anomaly detection, reducing the requirement for the number of defect samples.