Ampnet:An Advanced Lightweight Defect Detection Network for Tiny Steel Sheets Inside Mobile Phone in Industrial Scenarios

Abstract

In recent years, the rapid development of deep learning technology has provided many methods for defect detection in industrial application scenarios. However, defect detection in real-world industrial environments presents multiple challenges, including the difficulty of identifying tiny defects amidst complex backgrounds, textures, and other noise, as well as the inefficiencies and high error rates of traditional detection methods. To meet the detection requirements of modern industrial production, this study proposes a simulated industrial scenario where an industrial camera captures high-resolution images of small steel sheets under varying lighting conditions. Additionally, it introduces the Advanced Mobile Phone Steel Sheets Defect Detection Network (AMPNet). Firstly, this study presents a new attention mechanism (CASA) which can capture the extracted contextual information and enhance the central features, where the strip convolution can easily recognize and extract features of elongated defects. By integrating the spatial attention module (SAM) to enable the model to focus on the key information in the image more effectively. It can suppress irrelevant or redundant information, thus reducing the impact of noise on the model and improving the robustness of the model. Moreover, the CASA  attention mechanism and convolutional blocks are used to form a new module (RCS) by jumping connection, which not only alleviates the vanishing gradient problem, but also enables the network to capture more levels of features, improve the model's ability to learn and extract defect features, and does not significantly increase the amount of computation. This approach effectively addresses the challenge of extracting features from defects on small steel sheets with variable shapes, random positions, and extreme aspect ratios. Finally, by introducing the lightweight convolution module GhostConv, a new detection head has been proposed to significantly reduce the parameter amount and computational cost of the model. The findings reveal that AMPNet achieves a AP0.5 of 91.8\%, a Params of 3.0M, and a FLOPs of 5.5G, which greatly reduced the computing cost without increasing the number of parameters, and surpassing mainstream object detection models such as YOLOv5 and YOLOv8. This demonstrate that AMPNet has great recognition ability and high accuracy in industrial environments and is suitable for model deployment on resource-constrained embedded devices.

Keywords: Mobile phone Surface defect detection Lightweight Attention mechanism