Calculated Three-dimensional Reconstruction Of The Visual, The New Method Of Image Enhancement And Pattern Classification Problem

Computational vision is a science and technology that utilizes imaging devices to capture scenes and applies computers for analysis and processing. It aims to help or even substitute human to achieve vision tasks. Typical tasks in computational vision include, recognition, motion analysis, scene reconstruction, image enhancement, etc. Its applications in industry and human life include, fingerprint and face recognition, surveillance, 3D reconstruction, image deblurring, etc. This paper proposes new solutions to three important computational vision tasks with wide applications - 3D reconstruction, image enhancement and pattern classification.Major contribution of the paper includes, (1) For the 3D reconstruction, we propose to use a 'motion clue' - 'Relative Epipolar Motion(REM)' towards the correspondence problem in stereo vision. The proposed REM method is able to effectively establish feature correspondence given dynamic scenes with indistinguishable features. (2) For the image-deblurring task in the enhancement area, we argue that a dual(hybrid)-camera containing a high-resolution low-framerate camera and a low-resolution high-framerate camera is by far the first feasible solution to take advantage of multiple blur-kernels for deblurring, and we show from experiments that our solution achieves better performance over existing deblurring methods based on single kernel. (3) For the pattern classification problem, we propose a 'distribution concentration' property by introducing the 'Rectified Nearest Feature Line Segment (RNFLS)' classifier, which not only overcomes some drawbacks of the original NFL method, but also shows superior results on artificial and real-world classification tasks as compared to NN, k-NN, NFL, NNL methods.