Image Denoising Algorithm Inspired By Architecture And Population Coding Of Visual Cortex

With the comprehensive automation of industrial production and the rapid advancement of artificial intelligence technology,the field of image processing is experiencing unprecedented development opportunities.However,in practical applications,due to the influence of factors such as sensors,communication,and storage,images are often distorted by noise interference,affecting subsequent tasks such as image recognition and image segmentation.Therefore,image denoising,as one of the fundamental issues in digital image processing,plays a crucial role in improving image quality and utilization.Since 2014,research based on deep learning convolutional neural networks has shone in fields such as image recognition,segmentation,and denoising.Among them,residual structures(Res Net)and cross-layer connection structures(U-Net)have demonstrated widespread effectiveness in these image processing tasks.Although Res Net can alleviate training problems when the network has more layers,and U-Net acquires a strong multi-scale feature extraction capability by increasing the number of channels layer by layer in combination with pooling,there is an upper limit to the performance improvement brought by simply increasing network depth and channel numbers.On the other hand,while network modules or structures carefully designed by different researchers have shown their effectiveness,the stacking of these structures makes the network design increasingly complex and lacking in interpretability,making it more difficult for newcomers to make breakthroughs in the field of image denoising.Biological vision,after hundreds of millions of years of evolution,has a natural and powerful ability to ignore noise and carry out subsequent image processing.At the same time,the gradual unveiling of its principles by researchers makes it more interpretable compared to the black box of neural networks.This article is inspired by the encoding method of the brain’s visual cortex and carries out the following research work:First,this thesis attempts to extract an effective framework structure from the biological visual mechanism,integrating the columnar structure of the primary visual cortex and its encoding of visual information into the overall framework of the convolutional neural network.This structure adopts population coding and converts individual pixel values into multi-channel information based on specific selectivity.By preprocessing and expanding the number of channels,the feature learning pressure of the neural network is reduced,making it easier to capture the features of natural images.Next,this paper draws on the interactive relationship between orientation cortical columns and hue and brightness cortical columns.Machine learning is used to replace the manual assignment of orientation cortical columns,resulting in better edge extraction effects.The population encoding form of hue cortical columns can still retain the original information when the excitation level of the neuron population increases or decreases.This enables the network to use edge information to constrain the propagation of color information during the convolution iteration process,allowing the image to better preserve edge details and color purity on both sides of the edge during denoising.Finally,through objective quantification indicators and subjective visual effect verification,the visual cortex encoding framework used in this paper effectively improves the lower limit of the denoising network,and the suppression mechanism of edge information on color information further improves the denoising effect of the proposed model,making it one of the best denoising models.In addition,the model presented in this paper demonstrates strong robustness in noise tasks with non-training intensities.