Graphics And Image Processing Methods Based On Biological Visual Model

Researching the working mechanisms of biological visual systems, building corresponding machine visual systems and applying them to various graphics and image processing tasks are the frontier topics of artificial intelligence, computer vision, computer graphics and digital image process subjects. The signal processing mechanisms of the visual systems of human and mammals are accurate and complex, sensing the environment thoroughly in real time and then making decision rapidly. They are better than most machine visual systems in many aspects. So learning from the latest research results of visual neuron subject about visual cortex structure and information processing procedures, building and perfecting corresponding machine visual models and applying them to various graphics and image processing tasks have great scientific significance and application value.Pulse coupled neural network (PCNN) is a biological visual model which make reference to synchronous impulse phenomenon in cat visual cortex neurons. As a new model for digital image process, it shows the cluster feature for spatial proximity and intensity similar. It has been applied to a large range of image processing tasks like image filtering, image segmentation, image fusion, image edge and target detection, image feature extraction, etc. According to the analysis and summary of relevant research work in literature, on the one hand this paper extends PCNN from digital image processing field to computer graphics field, like point set surface denoising, mesh smoothing, merging and stitching, one the other hand this paper studies further on color image denoising, color image segmentation and color image fusion problems based on existing progresses. Our contributions are summarized as follows:Firstly, we proposed a method for point set surface denoising based on PCNN. The method mainly includes two steps:locating the noise points and smoothing the located noise points. Firstly, a pulse-coupled neural network for the point set surface is constructed. The stimulation value of each neuron is decided by the differences of the position and the normal of the K-nearest neighbor points. The noise points are located through the adaptive firing capture feature of PCNN. Based on the idea of bilateral filtering, the located noise points are smoothed, while the non-noise points remain their geometry position. Experimental results show that the proposed method not only removes the noise points efficiently, but also keeps the geometrical features of the models.Then we proposed a method for mesh smoothing, merging and stitching based on PCNN. Mesh is smoothed with PCNN method firstly. Via merging vertices in the overlapping regions, mesh stitching with large overlaps is achieved to avoid the filling-hole operations and small triangles produced by clipping edges. The overlapping regions are detected by oriented bounding box rapidly. With the method of moving least squares, the vertices in the overlapping regions are smoothed in order to reduce the noise produced by inaccurate alignment. Then, we merge the vertices with distance less than given tolerance and triangulate the remaining vertices with the constraint of common boundary edges. The final mesh is created by stitching the several meshes via the common boundary edges. The experimental results show that the proposed method can create merged mesh with good quality and the performance is acceptable in practice.Then we proposed a method for color image salt-pepper noise filtering based on PCNN. Firstly the method enhances the color image via histogram variance. Then it locates noise pixels in the image by utilizing PCNN’s specific feature that the fire of one neuron can capture firing of its adjacent neurons due to their spatial proximity and intensity similarity. Finally it removes the noise pixels by a VMF-likely vector filtering. Experimental results show that the proposed method is able to preserve fine details while removing salt-pepper noise.Then we proposed a method for color image segmentation based on PCNN. Firstly, the method enhances color image via Shannon entropy. Then a PCNN for the color image is constructed. Make the PCNN fires in iteration. The method uses the maximum variance ratio to decide the iteration count. Finally the method use Shannon entropy to select the output segmentation result from the iterations. The experimental results show that the proposed method performs well in both results and efficiency.At last, we proposed a method for color image fusion based on PCNN. Firstly, the method enhances color images via weighted histogram variance. Then it use PCNN as fusion operator for the contrast pyramid. The method is applied to multi-focus and multi-exposure images. The experimental results show that the proposed method performs well in the fusion tasks.