Yolov4 Algorithm for Target Detection in Existing Intelligent Waste Sorting Systems

Abstract

The problem of garbage sorting has caused a lot of trouble. Some target detection algorithms are being developed for the current situation of garbage sorting in China. In order to understand the current state of development, we investigate several algorithms suitable for object detection. This paper mainly analyzes three parts of yolov4 network model. In the backbone network, CSPDarkNet53 achieves dimensionality reduction through five convolution kernels, and MobileNetV3 replaces the activation function based on MobileNetV2 and uses MnasNet and Platform-Aware for optimization. Ghostbottleneck is obtained by superposition Ghost module, and then build GhostNet lightweight neural network.The principles of the three attention mechanisms are also explained，SENet using a single-channel convolution kernel，CBAM contains two plates Mc and Ms and CA has a length and width of the two channels. Finally, we explain a data enhancement where four random images are cropped and spliced together.