Research On Garbage Classification Method Based On Deep Learning

With the continuous advancement of urbanization in China,the production of household waste is also increasing year by year.Since the implementation of garbage classification standards,preliminary results have been achieved in garbage classification work.However,due to the wide variety of garbage types and difficulties in identification,garbage sorting workers have low efficiency in handling,and there is a certain risk of injury.The garbage image classification method based on deep learning can achieve fast and accurate garbage classification,improve resource recycling efficiency,and provide a foundation for the implementation of automated garbage classification.This article conducts research on deep learning based garbage classification methods to achieve accurate and fast garbage classification.The main research content of this article is as follows:(1)Garbage image classification based on weighted multi-scale feature fusion.For garbage image classification tasks with multiple categories,large intra class differences,and complex backgrounds,this article improves the Xception network by designing four feature extraction modules with different structures to study the different manifestations of structural differences in feature extraction and enhance the network’s ability to focus on garbage targets of various scales;combining coordinate attention and channel attention,obtain long-distance responses at the channel level and position space,and suppress the impact of irrelevant noise.In the experimental section,the performance of various feature extraction structures and different attention embedding methods were compared,and the combination module with the best comprehensive performance was selected as the feature extraction of the network to achieve accurate classification of garbage images.The final improved network’s classification accuracy on garbage datasets was 96.69%.(2)A garbage classification network based on lightweight CNN.In order to balance accuracy and parameter requirements,and considering the excellent performance of multi-scale convolutional stepwise concatenation structure in feature extraction,this paper selects the hollow concatenation structure as the basic feature extraction module.Drawing inspiration from the inverse residual structure design of MobileNetV2 network,pointwise convolution is used in the feature extraction module for feature dimensionality reduction,achieving lightweight operations.At the same time,a lightweight attention mechanism ECA is introduced to strengthen the network’s focus on garbage targets.In the end,on the constructed garbage dataset,compared to the basic network,the improved network improved its accuracy by 0.87%,and the model volume was only 0.44MB.The testing time on Raspberry Pi 4B was only about 26ms.(3)Design and implementation of a visualization system for household waste classification.This article designs a visualization system for household waste classification based on PyQt language.The system supports users to upload photos and provide feedback on recognition results,including information such as category and prediction confidence.Input garbage images with significant differences in scale,appearance,and background interference into the visualization system to test the effectiveness of the system and the robustness of the improved algorithm.The test results meet the expected requirements for garbage classification in practical situations.