轻量化的ResNet50图像分类模型OA

Aiming at the problem that the accuracy of residual network may be affected by the lack of feature extraction ability when dealing with image classification tasks,this paper proposes an EResNext50-EFPN image classification model.Firstly,the grouped convolution attention residual unit is designed,which uses an grouped convolution to replace the traditional convolution on the backbone path of the residual block,and incorporates the ECA attention mechanism,which not only enhances the expres-sion ability of the model,but also effectively reduces the amount of parameters and calculation.Secondly,we optimize the down-sampling module.Finally,a multi-feature fusion module is designed,which can effectively fuse features from different levels,so as to further improve the accuracy of image classification.In terms of model training,the warmup strategy is combined with the cosine annealing attenuation method to ensure that the model can converge more stably.Experimental results show that compared with the original ResNet50 model,the classification accuracy of EResNext50-EFPN model is improved by 2.95 percentage points on the CIFAR-100 dataset,while the number of parameters is only 69%of the ResNet50 model,and the calculation amount is reduced to 60%of the ResNet50 model.