Lightweighting of Deep Learning Models with Application in Object Detection

In the field of industrial manufacturing, the deployment and operation of models on resource-constrained devices (e.g., mobile devices, embedded systems) is of paramount importance. To achieve lightweight target detection models, we propose an improved model based on YOLOv5s. This method optimizes the backbone feature extraction network using MobileNetv3 modules and enhances the model neck with GSConv convolutions. Furthermore, to address the performance degradation caused by model lightweighting, we integrate a Coordinate Attention (CA) mechanism into the architecture. The resulting model is denoted as YOLOv5s-MV3_B. To validate the model's efficacy, we conduct experiments on a object detection task. Experimental results demonstrate that the lightweight model achieves a 61.39% reduction in overall computational complexity and a 48.25% compression in model size compared to the baseline YOLOv5s.