Garbage Recognition Based On Lightweight Neural Network

As people relentlessly pursue materialistic life and the living standard continues to improve,the amount of waste produced in daily life is increasing rapidly and constantly affecting the ecological environment we rely on for survival.Therefore,it is very important to correctly understand garbage classification and conduct effective garbage classification and treatment.In order to address the issues of fast detection of targets and lightweight deployment of models while maintaining accuracy,this article uses computer vision technology based on deep learning to compare the comprehensive performance of Two-stage Faster region convolutional neural networks(Faster R-CNN),One-stage Single Shot Multi Box Detector(SSD)and You Only Look Once v3(YOLOv3)in identifying household waste.Based on the results,YOLOv3 was chosen as the base network for lightweight processing,and MobileNetV3,ShuffleNetV2 and Ghostnet were selected to replace the original feature extraction network to optimize the network structure.The results showed that when MobileNetV3 was used to replace the original Dark Net-53 feature extraction network of YOLOv3,its combined features of deep convolution and pointwise convolution allowed for a significant reduction of65.2% in network size and about 65.6% in computing parameters while improving recognition speed by 27.5%,even though its recognition accuracy decreased slightly.ShuffleNetV2 utilized information recombination among different channels,which increased the fusion communication between channels but sacrificed some accuracy.Compared with the original YOLOv3 network,the accuracy decreased by about 2.7%,but the model size was reduced by 62.8%,and the parameter amount decreased by64.9%.When Ghostnet was used as the backbone network for feature extraction,the model was compressed by about 47.5%,the parameter amount reduced by nearly50.6%,and the model complexity decreased by about 61.7%.Although it was slightly inferior to MobileNetV3 and ShuffleNetV2 in terms of model compression and parameter computation,Ghostnet-based network lightweight processing still maintained an accuracy rate of up to 96% due to its ability to maintain the information of relevant feature maps through mapping and reduce the loss of feature extraction information.Finally,by lightweight processing of the Ghostnet feature extraction network,system design was carried out using the Nvidia Jetson TX2 embedded development board,and a Python-human-computer interaction interface was implemented using PyQt to achieve correct identification and classification of household waste.After the model is accelerated by TensorRT,the inference time of the model is improved from17.99 frames per second to 33.33 frames per second on the premise of maintaining the quasi-high accuracy,which greatly improves the reasoning rate of the model and has strong practical application significance.