Image Classification Algorithm Based on Improved Soft Thresholding and Residual Network

To solve the problems in information loss, insufficient feature extraction, and overfitting of the ResNet model in image classification tasks, this paper proposes an improved image classification algorithm (ST-ResNet18). By changing the Conv1 block and delaying the downsampling operation in the residual module, this algorithm solves the problem that the residual network loses input features in the feature extraction process. Meantime, a soft thresholding operation is added to the residual module to improve the expression ability of the feature map for the enhancement of the classification performance. In addition, the erasure data augmentation technique is used to improve the algorithm generalization ability and avoid the overfitting problem. The experimental results suggest the accuracy of the proposed algorithm on Cifar10 and Cifar100 datasets reaches 96.14% and 79.85% respectively, which is 1.39% and 3.12% higher than that of the original image classification algorithm.