Research on target defect detection method based on Canny-YOLOv7 improved model

• 1、Artificial Intelligence Security Innovation Team, Beijing Information Science and Technology University, Beijing 100192, China
• 2、Department of Information Security, Beijing Information Science and Technology University, Beijing 100192, China

Abstract：To address the problem of insufficient defect detection capabilities in current military and industrial products, we propose a fusion enhancement model based on the Canny-YOLOv7 algorithm, focusing on small target defect detection common in production. The model primarily utilizes the Canny edge detection algorithm to identify potential defect areas in images, enhancing the features of these areas. Building on this, we improve the YOLOv7 model to enhance its generalization ability. First, to improve the model\'s capability to extract subtle defect features, we replace the ELAN module in the original backbone network with the Swin Transformer (STR) module. Second, we incorporate the SIOU loss function, enabling the improved model to quickly learn accurate target positioning. Finally, the LeakyReLU activation function is adopted to reduce computational overhead and further improve the model\'s detection speed. Experimental results demonstrate that, under the same dataset, the proposed enhanced model achieves an average detection accuracy of 97.5%, which is 4.6% higher than the original YOLOv7 model. Additionally, the detection speed （FPS） reaches 52.45, meeting real-time detection requirements.

Keywords： deep learning defect detection YOLOv7 model Canny algorithm Swin Transformer