CPDD-CLMM: a comprehensive lightweight mobile-optimized network for composite plate defect detection.

Automatic defect-detection technology based on deep learning is increasingly used for distinguishing production quality by many industries. However, production lines are usually installed with lots of function modules, which make it difficult to integrate new modules. Common deep learning models run on PC platforms and require a big space with high cost, while ARM64 mobile platforms are much smaller with less cost and equivalent connectivity but also weaker performance. Considering these facts, ARM64 platforms with a fully optimized model are the best solution for adding a defect-detection function for existing production lines. This paper focused on a mobile-optimized model to achieve higher speed and equivalent precision on the ARM64 mobile platform for detection. First, the model structure is simplified by reducing the redundancy of feature maps to increase the network inference speed. Second, a convolutional block attention module is attached to compensate for the decrease in precision caused by structure simplification. Furthermore, a transfer learning method is adopted to improve training performance. Finally, the trained and compiled module is exported to the PyTorch Mobile format and deployed on the mobile platform application to execute its defect-detection function. The results show that the optimized network achieves a speed of 2.124 fps, 210.7% compared with that of You Only Look Once v5n, i.e., 1.008 fps, on the RK3399 ARM64 platform, and has an average mAP of 99.2%. The studied mobile-optimized model has better speed and equivalent precision and can be available on many different ARM64 platforms regardless of the processor manufacturer. It can satisfy the need for real-time defect detection and can be used in similar scenarios. In the future, more improvements could be made such as deploying on platforms with NPU support to achieve faster speed, exploring the relationships between dataset properties and transfer learning effects, even training and running the model directly on ARM64 platforms. [ABSTRACT FROM AUTHOR]