| With the development of computer vision algorithm and and the increasing comptation capabilities,many vision based applications have been introduced into our daily life.However,most of the systems base on the two-dimension color images,which are easily suffered from noise,luminance changes,etc.,which will do harm to the system’s reliability and robustness.This paper aims at improving the vision perception ability through different way of methods.Specifically,we study how to improve the vision perception ability through image restoration/enhancement,information fusion,discovering geometric information from images.The main research contents and their corresponding novelties are as follows:Firstly,we present a novel method to remove rain streaks in videos based on rainy intensity proposed.The rainy intensity is defined according to the positive intensity fluctuations and intensity symmetric along the pixels time evolution.The rainy intensity measures the rains heavy level very well,and helps to estimate the pixels original value in highly rainy scene.we design a hardware accelerator implemented by FPGA for single image haze removal based on dark channel prior and guided filter.We modify the calculation of the dark channel and atmospheric light,so as to make the algorithm be implemented in hardware easier and reduce the computation greatly while making almost no difference to the original algorithm.Fast mean filtering and pipeline technique are applied to speed up the processing procedure.The hardware accelerator can process single image haze removal is able to process about 80 haze images(720 × 576)per second.It meets the requirement of real time processing.Secondly,we present a fusion framework for background subtraction based on color and depth data.Background subtraction algorithms based on purely color or depth data are easily suffering from performance degradation in some challenging situations such as shadowing,color camouflage and depth camouflage.The foreground segmented will be very inaccurate.We formulate the fusion task as a classification problem to improve the background subtraction results.We propose some discriminative features that are extracted from the source color and depth data considering spatial and temporal information,to help improve the results.Experiments show that our method can take full advantage of the both information to detect foreground in color camouflage,shadowing and depth camouflage situations,being robust with different kinds of background subtraction algorithms in different scenes,and outperforming the state-of-the-art algorithms that fuse color and depth data for background subtraction.Thirdly,we leveraged the geometry prior knowledge of stereo matching,i.e.,the reference image can be reconstructed from the second image in the visible regions using the predicted disparity for learning a CNN to estimate disparity.The network is trained with the sparse ground truth disparity in a supervised learning manner,and trained to satisfy the geometry prior in a self-supervised learning manner.The geometry prior is useful especially when the ground truth disparity is sparse(for example,the outdoor scene of KITTI 2015),and help the network to perform better.Besides,we propose a network architecture to incorporate all steps of stereo matching.The network can be roughly divided into two sub-networks.The first sub-network performs matching cost calculation,matching cost aggregation and disparity calculation to produce the initial disparity.The second sub-network uses the feature constancy of the initial disparity to refines the initial disparity through a Bayesian inference process.To calculate feature constancy,we introduce construction error and correlation in feature space to efficiently identify the quality of the initial disparity.The proposed method has been evaluated on the Scene Flow and KITTI datasets,and achieves the state-of-the-art performance on the KITTI 2012,KITTI2015 benchmarks while maintaining a very fast running time. |
PDF Full Text Request