An improved denoising model for convolutional neural network

Abstract

In recent years, computer vision and many other research fields have put forward higher and higher requirements for image quality, and image denoising has become an important research direction. Aiming at the problems of general image denoising effect and long algorithm convergence time, this paper makes an improvement on the original network structure of denoising convolutional neural network DnCNN model and adds channel attention mechanism in the penultimate layer. The channel attention mechanism further improves the signal-to-noise ratio (SNR) and optimizes the noise reduction effect through convolution and deconvolution neural networks. The experimental results show that, under the same noise reduction level of 25, the PSNR of the proposed DnCNN2 algorithm is 1.25dB and 1.62dB higher than that of the DnCNN algorithm after the verification of the BSD68 and SET12 test data sets. Structural similarity (SSLM) was increased by 0.0011 and 0.0103, respectively. Meanwhile, the convergence time of the algorithm becomes shorter. Compared with traditional deep learning image denoising algorithms, the noise reduction effect and operation efficiency of the proposed method are also improved, which has certain advantages.