Research On Garbage Image Classification Method Based On Improved Swin Transformer

With the continuous development of the social economy,the production of garbage in daily life has rapidly increased,and the requirements for garbage classification and treatment are also increasing.However,the efficiency of manual garbage classification is very low.In environments where a large amount of garbage is processed,we need to use computer image recognition classification technology to improve the efficiency and accuracy of garbage classification.Aiming at the problems existing in traditional image classification,such as insufficient extraction of image category features,limited number of data samples,limitations of multi-target recognition,and interference of complex environment to image recognition and classification,this paper conducts targeted research.Firstly,research is conducted on four common classic image classification algorithms,and preprocessing operations such as data enhancement and normalization are performed on the image dataset to be classified.Through experimental comparison,the Swin Transformer algorithm with the highest classification performance is selected as the benchmark model from the four image classification models.Because the benchmark model did not add bias terms when calculating attention,it was unable to supplement the original information.After calculating attention,it only had an attention enhancing effect,which would result in losing some of the original information.There are two improvements and optimizations made to the benchmark model.Firstly,the original two parts in the Swin Transformer: dividing the input image into multiple small parts,Patch Partition,and mapping each small part to a low dimensional inward representation.Linear Embedding has been replaced with a Convolutional Token Embedding module for feature extraction and encoding through convolutional operations.The second is to optimize the Swin Transformer Block by using a windowing approach to divide the input image into multiple blocks and perform multi head self attention operations on each block.The experimental results show that the optimized Swin Transformer model has improved the accuracy of image classification,thus proving the effectiveness of the improved model.