Research on object detection algorithm based on improved YOLOv7

Abstract

Aiming at the problems of low detection speed, high leakage rate, and reduced detection speed after migrating to embedded devices in the workshop object detection task, an improved object detection algorithm for YOLOv7 is proposed. In the improved algorithm, a lightweight ShuffleNetv2 network is introduced to improve the YOLOv7 algorithm based on the YOLOv7 network framework. Since the ReLU function is used in the ShuffleNetv2 network, and the ReLU function will cause neuron "necrosis" during the training process, the SiLU activation function in the YOLOv7 network structure is used to replace the ReLU function in ShuffleNetv2, which meets the accuracy rate and at the same time has a better performance. Therefore, the SiLU activation function in the YOLOv7 network structure is used to replace the ReLU function in ShuffleNetv2, which can satisfy the accuracy and at the same time have better detection speed. The experimental results show that the accuracy of the improved algorithm on the test set Precision (P) reaches 91.7%, Recall (R) reaches 81.4%, and mean Average Precision (mAP) increases by 2.9 percentage points. The improved algorithm can effectively guarantee the detection speed in the real scene detection task, and at the same time reduce the leakage rate to a great extent, showing good detection performance.