Computational Models And Applications Of The Retinal Color Vision

Vision,is a subjective perception of the objective world projected on the retina through the visual system processing.The ultimate aim of vision research is elucidating how the visual system produces this perception.It is not only conducive to comprehending the visual system mechanism and enhancing our understanding of human beings,but also beneficial to the technology improvement of image acquisition and processing to provide the input for the artificial intelligence machine.From the perspective of computational model,this dissertation,which is in the form of the physiological-based,taskoriented,computable and simplest system,discusses the possible retinal mechanisms of visual function such as color constant,image enhancement and brightness dynamic range compression,and applies them to relevant tasks.Introducing the feedback from horizontal cells to cones and the disinhibition effect of the non-classical receptive field of retinal ganglion cells into the computer vision task of color constancy,this dissertation proposes a non-illuminant-estimating color invariant computational model by imitating the highly correlated parts of the retinal mechanism,such as the feedback from horizontal cells,center-surround color oppoency receptive field of retinal ganglion cells,the disinhibition effect from the subunits of the non-classical receptive field,and the adaptation in ganglion cells.In the proposed model,the horizontal cells modulation provides a global color correction with cone-specific lateral gain control,and the ganglion cells refine the processing with the spatially sensitivity from the disinhibition effect of the non-classical receptive field.The adaptation in ganglion cells determines the key parameter of inhibitory weight automatically and provides adaptability for different environments.The competitive results in comparison with the state-of-the-art methods indicate that the feedback from horizontal cells and the disinhibition effect of the non-classical receptive field of retinal ganglion cells play an important role in color constancy of biological vision.In regard to the atmospheric scattering mechanisms,this dissertation proposes a nonpriori-assumption model for enhancing visibility of hazy images by imitating the highly correlated parts of the retinal mechanism,such as the difference of gaussians receptive field of bipolar cells,the adjustment from amacrine cells,the disinhibition effect of retinal ganglion cells and the ON/OFF channels.The proposed model,which uses the difference of gaussians receptive field of bipolar cells to remove the additive air light,adjustment from amacrine cells to enhance contrast and to compensate the direct attenuation,and the disinhibition effect of ganglion cells to recover the multiplicative distortion of the scene radiance and to enhance details,provides a simple fast algorithm for single image dehaze enhancement.The competitive results in comparison with the state-of-the-art methods indicate that bipolar cells,amacrine cells and ganglion cells contribute to signal enhancement in the visual system.Taking the light-induced changes in the size of horizontal cells’ receptive field into account,this dissertation proposes a tone mapping model for high dynamic range image rendering,by imitating the highly correlated parts of the retinal mechanism,such as the dynamic receptive field of horizontal cells and the difference of gaussians receptive field of bipolar cells.The key novelty of the proposed model lies in the adaptive adjustment of the receptive field size of horizontal cells based on the local brightness,which simulates the dynamic gap junction between horizontal cells.This enables the brightness of distinct regions to be recovered into clearly visible ranges while reducing halo artifacts common to other methods.Bipolar cells serve to enhance the local contrast with their centersurround RF structure.Corresponding to the amazing ability of the visual system to quickly achieve stable perception under varying natural light environments,our method provides a robust performance for various scenes.The competitive results in comparison with the state-of-the-art methods indicate that the feedback from horizontal cells and the difference of gaussians receptive field of bipolar cells,which works like a band pass filter,compress the dynamic range of the input signal effectively in favor of the fact that the vision system handles the input signals with limited resources.This dissertation,by discussing three different computational models,investigates the possible retinal mechanisms of color constant,image enhancement and brightness dynamic range compression respectively.The next work is to integrate them into a general retinal model for multi-vision tasks.