Research On Garbage Classification Method Based On Improved YOLOX Algorith

In recent years,China has attached increasing importance to the classification and processing of daily household waste,and major cities have implemented relevant policies on waste sorting.However,existing waste classification algorithms suffer from high network computational and parameter costs,making it difficult to deploy them on mobile devices.Therefore,this paper proposes a lightweight approach based on improved YOLOX.The aim is to reduce the model’s parameter size and computational requirements to facilitate its deployment on mobile devices.Through detailed experimental comparisons and result analysis,the effectiveness of the proposed method is demonstrated.The main research content is as follows:(1)Proposed improved lightweight method based on YOLOX model.Firstly,replaced the original IOU loss function with EIOU to overcome the problem of imbalanced positive and negative samples and improve detection accuracy.Secondly,introduced the CBAM attention mechanism in the neck network to redistribute weights across different channels,capturing both fine-grained features from shallow layers and semantic information from deep layers.Lastly,replaced the CSP module in the feature extraction network with the Ghost Bottleneck module to preserve more edge information and reduce model parameter count.Experimental results showed a12% reduction in parameters and a 0.3% improvement in accuracy.(2)Proposed an improved method based on depthwise separable convolution.Firstly,replaced the CSP structure in the neck network with lightweight modules and replaced the regular convolutions with depthwise separable convolutions to reduce the parameter count and computational complexity.Secondly,decoupled the detection head and replaced the two regular convolutions with depthwise separable convolutions.Lastly,introduced an improved CA attention module in the backbone network.Experimental results showed that the improved model achieved a 37%reduction in parameters and a 40% reduction in computational complexity without sacrificing detection accuracy.