Research On Visual Neuron Network Model For Image Processing Applications

The biological vision system has extraordinary ability to deal with external information.It is of great significance to study the visual perception mechanism for constructing an efficient visual neural computing model and applying it to image processing tasks.This paper deeply studies the application of different visual neural computing models in contour detection and segmentation tasks.On the one hand,we consider the contact between the retina,the lateral geniculate nucleus and the primary visual cortex.From the overall perspective,the contour detection model based on visual perception mechanism of visual pathway is constructed.On the other hand,we propose a contour detection model based on convolutional neural network to simulate the brain neural network and construct a pyramid feature decoding module to effectively improve the contour detection.We transfer our convolutional neural network to the specific liver tumor segmentation and the results demonstrate that our model can also solve the segmentation tasks.The main research contents of this paper are as follows:(1)A new contour detection method based on visual perception mechanism is proposed,which considers the visual information transmission and processing in the primary visual pathway.Firstly,the Gaussian derivative function is used to extract the primary contours.Then,the temporal and spatial coding of neuron information is used to improve the contour contrast.In addition,proposing a neural coding that can strengthen the redundancy of visual information in the optic radiation area to enhance the contour information and robustness of detecting.Finally,primary contour is fed forward to the primary visual cortex to achieve rapid adjustment of the contour response.When evaluating on the RuG40datasets,we achieve the optimal parameter for whole dataset 0.48 and each image 0.55 respectively,and speed FPS reaches 0.5.The method is more effective and faster than other algorithms.(2)Low-to-high hierarchical convolutional features can significantly improve contour detection.We propose a feature decoder-based algorithm that employs a Feature Decoder Network(FDN)to extract more information within limited Convolutional Neural Network(CNN)features.The feature decoder fuses CNN features of adjacent layers to judge the edge and non-edge pixels,which can learn the relationship and distinction between low-level edge hierarchies and high-level semantic hierarchies.Furthemiore,we use Gaussian blur labels to train the network to optimize network convergence and training.From the experimental results,our proposed algorithm perfomas better on the BSDS500(AP of 0.865)and NYUD(OIS F-measure of 0.775)datasets compared to the state-of－the-art algorithms,including RCF.(3)The experiment of deep convolutional neural computation model in medical image processing is carried out.Based on the LiTS-Liver Tumor Segmentation Challenge,We transfer our method FDN to the specific liver tumor segmentation.The sigraod activation function is replaced with the softmax activation function,and the weighted cross entropy function is used to ease the problem of imbalance segmentation targets.The results show that our model is superior to the other models in this segmentation task.