The Lightweight Improvement of YOLOv7-tiny and its Application on Defect Detection

Addressing the need for visual defect detection in the manufacturing process of certain vehicular ultrasonic radars, this paper introduces an enhanced algorithm for the YOLOv7-tiny object detection model to ensure real-time and accurate detection. The modification involves replacing the YOLOv7-tiny's backbone network with a more lightweight FasterNet-T0 module, aiming to reduce the model's parameter count and computational demand, thus accelerating detection speed. To better recognize defects in components with slender structural features, a Dynamic Snake Convolution (DSConv) is utilized in place of the traditional convolution algorithm in the detection head, enhancing the model's defect identification capabilities. Concurrently, the adoption of a more versatile loss function, Alpha-IoU, as the evaluation standard, improves the accuracy of bounding box predictions. Experimental tests on sample products reveal that, compared to the original YOLOv7-tiny model, our approach reduces the parameter count by 27.6% and floating-point operations by 34.4%, while elevating the mAP@0.5 metric from 96.5% to 97.8%. This improvement in detection precision for specific targets, along with the model's light-weighting, makes it suitable for defect.