Improved Deep Residual Shrinkage Network for Noisy Image Recognition

In order to enhance the recognition capability of noisy images more effectively, this paper proposes a novel network model based on the improvement of deep residual shrinkage network. A new approach for obtaining soft thresholds is introduced. This improved soft threshold is inserted as a non-linear transformation into deep networks, thereby eliminating insignificant information features in the images. Moreover, this soft threshold can be combined with neural networks for joint training, automatically determining the threshold magnitude. We experimentally validate the proposed method using a manually augmented noisy Flowers dataset, demonstrating its effectiveness in improving image recognition accuracy. Additionally, compared to the deep residual shrinkage network model, it exhibits a smaller parameter count.