Research On Garbage Image Classification Based On Image Fusion And Attention Mechanism

With the rapid development of social economy and culture,the amount of domestic waste produced by people every day is also increasing.Therefore,the recyclability of domestic waste plays a crucial role in modern society,helping to reduce various types of pollution and bring economic benefits.To achieve this goal,garbage classification is one of the most important steps in the recycling process.Although people have already classified their garbage when they put it in,due to the limited knowledge of garbage classification and the current use of container level classification for garbage,manual classification in the later stage poses problems of danger,complexity,and low efficiency.Therefore,there is a need for a garbage classification algorithm to guide people when placing garbage to avoid placing errors.The current popular deep learning technology can provide high performance visual recognition model,which is conducive to the automation of garbage sorting task.However,there are three challenges in using depth recognition model directly,that is,lack of sufficient data and high noise of data quality;In the aspect of image feature extraction,the surface features of different types of garbage images have little difference,which is easy to be confused.In terms of image classification,two kinds of garbage are similar in many aspects,which makes it difficult for Softmax loss function to classify these two kinds of garbage,which is prone to misjudgment.So we propose a new network model to address these challenges.The main work of this paper is as follows:As for the selection of the basic model for garbage classification,Res Net-34 was finally selected as the basic configuration for garbage classification in this article through analysis of the advantages and disadvantages of the network model structures of the five most popular algorithms and experimental comparison.At the same time,this article also conducts data analysis on selected public datasets(Huawei Garbage Classification Challenge Cup)and self-built datasets(Garbage),and performs pre-processing such as filling,scaling,cropping,and data enhancement to improve the resolution of garbage images.In order to improve the classification ability of the model,a new Res Net-34 network model is proposed,which is mainly optimized in three parts.Multi-feature fusion of input images is adopted,and three parallel branches are adopted in the input part,that is,convolution kernels of different sizes are used for feature extraction;After the residual network structure module,an attention mechanism is added,which combines the constructed Gram matrix,average pooling,and maximum pooling to make the model focus on the parts that are conducive to classification and achieve better feature extraction functions;Use Softmax loss and Center loss functions to optimize network parameters,making the feature information extracted from the model have better generalization and discrimination capabilities.Finally,this paper compares the proposed classification model with other garbage classification methods,verifying the advantages of the proposed algorithm,and visually demonstrating the classification effect of the model.