Lightweight Improved Based on YOLOv4 Object Detection Algorithm

Abstract

To address the problem that the existing object detection network models are large in size and complex in operation and cannot satisfy both detection speed and accuracy under the limited resources and small size platform. Based on YOLOv4 as the benchmark network, a lightweight object detection model LW-YOLO is proposed. Firstly, the backbone feature extraction network is replaced with MobileNetv1, while the number of feature fusion network parameters is significantly reduced by the depth separable convolutional module. Then the BN layer coefficients are used as scaling factors for the importance of the convolutional channels, the scaling factors are sparse using polarization regularization, the errors before and after pruning are reconstructed using least squares and channel weighting methods. The appropriate pruning thresholds are obtained by minimizing the reconstructed errors, the channels with small scaling factor values are eliminated to achieve the lightweight. The experimental results on the VOC (Visual Object Classes) dataset show that the detection accuracy of LW-YOLO is 87.00%, and the FPS(Frames Per Second ) reaches 48.89, which is better than the original YOLOv4 algorithm. It also significantly reduces the number of parameters, computation, and model size, which is more suitable for application in resource-poor embedded mobile devices.