Hybrid Garbage Classification Based On Convolutional Neural Networks

With the improvement of people’s living standards,the issue of garbage classification and garbage disposal is receiving increasing attention from society.At present,many garbage classification target detection models and applications have emerged in the field of garbage classification and recognition.However,there is still great room for improvement due to the common problems of poor real-time performance,limited detectable garbage categories,inability to detect occluded objects,and the ability to detect only a single item at a time.This thesis proposes a hybrid garbage classification method based on convolutional neural networks,which not only ensures reasonable real-time performance but also significantly improves detection performance.The main work of this thesis is as follows:1.In the selection of mixed garbage classification models,the Faster R-CNN model and YOLOv7 model were selected for comparative experiments on two-stage object detection technology and single-stage object detection technology in the field of object detection.The garbage classification dataset used in the experiment consists of 4 major categories and 44 sub categories,totaling 1 5064 pieces.By comparing and analyzing the average detection accuracy and detection speed of the two models on all garbage categories,YOLOv7,which is more real-time,was ultimately selected as the improved basic model in the experiment.2.In order to enhance the network’s ability to fit nonlinear functions and improve the overall average detection accuracy of the model,this thesis compared the training effects of different activation functions on the model through comparative experiments,and selected SiLU as the activation function of the model.For the gradient descent optimization problem of the model,this work chose multiple gradient descent optimization algorithms for comparative experiments.The experiments suggested SGD random gradient descent as the gradient descent optimizer of the model,and corresponding modifications made based on the training situation of the dataset on the model.In order to reduce the IOU loss between the prediction frame and the real frame in target detection,different frame loss functions were used in the model for training,and finally SIOU Loss was decided as the model’s frame loss function,which effectively improves the model performance.In order to improve the model’s ability to extract features and effectively detect garbage in complex scenes and garbage with mutual occlusion,this work studied the introduction of lightweight attention mechanism into the backbone network of the model,and made an experimental analysis of lightweight SE attention mechanism and SimAM non-parametric attention mechanism.Experiments showed that the SE attention mechanism enables the model to better learn image features and improves the average detection accuracy of the model on various garbage categories.3.Introducing the above four mentioned improvements into the YOLOv7 model yields the final hybrid garbage classification and detection model YOLOv7-SOIE.Comparing the training results of YOLOv7-SOIE with those of Faster R-CNN and YOLOv7,it is found that YOLOv7-SOIE has improved detection accuracy on the majority of the garbage categories,an increase of 2.1%with average detection accuracy mAP@.50 compared to YOLOv7 and 14.8%compared to Faster R-CNN.Although the computational complexity of the proposed model increases a little,its real-time performance is still satisfactory.