Research on lightweight YOLO visual detection method for complex scenes

Abstract

The existing deep learning object detection algorithms (such as the native YOLO series) often encounter problems such as high computational resource consumption, high inference latency on embedded devices, and decreased detection accuracy in extreme environments when dealing with complex visual scenes (such as drastic changes in lighting, severe occlusion, and background noise interference). To address these issues, this paper proposes a lightweight YOLO detection method for complex scenes. It includes a lightweight backbone network based on feature reuse, which can construct a "multi-scale fusion - attention enhancement" detection system on the dataset and uses knowledge distillation technology to optimize the model's generalization ability on low-power platforms. Experimental results show that the light weight model provided in this paper has higher average detection accuracy compared to the original YOLOv8; the model provided in this paper has reduced parameter quantity by 42.6% compared to the original; and the inference speed of the model on the embedded edge computing platform has increased by 31.2%. Therefore, this provides a very efficient and accurate visual perception solution for complex industrial environments, mobile visual tasks, etc.