Research On Solid Waste Target Detection Method Based On Lightweight Deep Neural Network

Machine vision has achieved ideal results by utilizing visual sensors to obtain image information and utilizing convolutional neural networks for feature extraction and classification.Nowadays,the large amount of solid waste generated by the retirement of old nuclear facilities urgently needs to be treated.However,it’s very dangerous for manual operations in such hazardous working environments.Using machine vision in conjunction with relevant hardware equipment to replace human labor can greatly improve work efficiency and reduce personnel contact.However,some mainstream object detection networks in the field of machine vision today have high computational complexity and multiple parameters,which makes it difficult for these methods to maintain fast detection speed on mobile hardware platforms with limited computing power.This article focuses on the current requirements for target detection and classification in solid waste disposal processes,and conducts research on lightweight target detection networks.The backbone and bottleneck parts of the YOLOv5 target detection network are modified to reduce the number of model parameters and improve model detection speed while ensuring model recognition accuracy.The main research work is as follows:(1)Making lightweight transformation based on YOLOv5 object detection model.This article proposes a lightweight object detection model named YOLOv5-ms,designs an M-block structure to transform the YOLOv5 backbone network,and uses Sim SPPF spatial pyramid pooling to replace SPPF to improve computational speed.Compared to the benchmark model,the parameter count of YOLOv5-ms decreases by 28% and the computational complexity decreases by 27%;(2)Proposing a YOLOv5-ms Att object detection model that integrates attention mechanism.Further improvements were made to the bottleneck part of the YOLOv5-ms model,which includes incorporating the Sim AM attention mechanism,and using the GSConv convolution module to replace standard convolution.While achieving lightweight and improved detection speed,the accuracy of YOLOv5-ms Att was improved by 0.7 percentage points and 1.0percentage points compared to YOLOv5 s and YOLOv5-ms respectively;(3)Designing and implementing a lightweight object detection software system.This article develops a graphical human-computer interaction interface that combines with the YOLOv5-ms Att algorithm to form a complete object detection software system,achieving real-time detection of images or videos with different inputs.This project proposes a lightweight object detection algorithm named YOLOv5-ms Att,and designs and implements a lightweight object detection software system which can identify and classify targets accurately and quickly in solid waste sorting scenarios,which has practical application value in promoting the efficient application of robots in solid waste disposal work.