HL-Net: Zero-shot image denoising algorithm based on hybrid learning

Deep neural networks trained on large datasets have achieved good results in image denoising. However, networks trained on specific datasets often have poor generalization, which is not conducive to practical applications. In addition, most denoisers often ignore the detail differences between image texture areas and smooth areas during the denoising process, resulting in the loss of details in texture areas. In this paper, we propose a zero-shot denoising network (HL-Net) with a Texture Noise Attention module (TNA) and a hybrid learning architecture. TNA allows the network to better extract noise information from texture areas without destroying their details, while the hybrid learning architecture uses two denoising methods for image smooth areas and texture areas: content learning LC (learning non-noise information in noisy images) and noise learning LN (learning noise information in noisy images), which can better preserve the image details in complex texture areas while extracting content information from smooth areas of noisy images. Additionally, the loss function proposed for the zero-sample hybrid learning network uses residual loss, consistency loss, and mean loss, enabling the network to better perform content learning and noise learning. Experiments show that our method outperforms most zero-sample denoising methods on various types of synthetic noise datasets and real-world noise datasets.