An efficient detection method for surface defects of Grosvenori based on improved YOLOv7-tiny

In this paper, we present an enhanced YOLOv7-tiny network model aimed at improving appearance defect detection in the Momordica grosvenori production process in Guangxi. Our model addresses the challenges of low accuracy in existing object detection algorithms and the stringent real-time requirements of automatic sorting systems. We introduce a novel complex attention mechanism (CAM module) to enhance the network's feature extraction capabilities, thereby increasing detection accuracy. By incorporating depthwise separable convolution into the backbone network, we achieve a lightweight model with reduced parameters and computational load. Additionally, we optimize the SPP structure for better adaptability to varying feature scales. The Focal Loss function is utilized to prioritize hard-to-classify samples, further boosting overall detection performance. Experimental results indicate that our algorithm achieves a detection accuracy of 96.73% and a speed of 60.1 frames per second on a self-built dataset, outperforming other algorithms and fulfilling the requirements for real-time defect detection with high accuracy.