Research On Image Classification Method Of Domestic Garbage

As people’s consciousness of protection of the environment and awareness of resource recycling increases,garbage classification has gradually become a hot topic.In comparison to manual classification methods,the garbage classification technology using convolutional neural networks has high efficiency.However,garbage exhibits diverse shapes,sizes,and application scenarios,posing challenges in feature extraction and real-time interaction for garbage classification models.Therefore,the main focus of this thesis is to construct a garbage image classification algorithm with high accuracy and relatively fewer model parameters.The main contributions and innovations are as follows:(1)In response to the complex environmental backgrounds and large intra-class variances of garbage,we propose a garbage classification network structure that combines Efficient Net V2 and an improved attention mechanism.The improved attention mechanism,called GSENet,integrates salient local features and detailed information in the image,enabling the network to focus on important features in the feature map,thereby enhancing the network’s feature extraction capability and suppressing the influence of background information.To integrate multi-scale information and address the significant variation in garbage sizes,we design the E-SPP structure.To tackle the problem of imbalanced class distribution,we adopt data augmentation strategies and introduce the Focal Loss function during training.Experimental results demonstrate that the improved network achieves an accuracy of94.3%,which is a 4.6% improvement over the original network.(2)To address the issues of low garbage detection accuracy,missed detections,and large network parameters,we propose an end-to-end garbage object detection algorithm called GYOLOv7.This algorithm replaces the base network in YOLOv7 with the improved Efficient Net V2 to ensure the network’s ability to extract garbage features.Furthermore,we improve the feature pyramid network using the Slim-neck structure,which combines standard convolution and depthwise separable convolution information while preserving hidden connections between each channel.This effectively improves the model’s detection accuracy and reduces the number of model parameters.Finally,to reduce the missed detection rate in dense regions,we introduce the DIo U-NMS algorithm during the post-processing stage of frame detection.Experimental results demonstrate that the GYOLOv7 network achieves a 2.53%improvement in detection accuracy(m AP)compared to the original network,reaching92.5%,while reducing the number of parameters by 13.1M.Besides,a garbage classification image platform on the basis of GYOLOv7 is designed and implemented.The main functions of the garbage classification interface include detection,classification.