Research on High Performance Edge-Preserving Image Filtering Algorithm for Stereo Matching and Image Dehazing

Edge-preserving smoothing operator, such as the guided image filter, is designed to effectively preserve the structural information during image filtering. It is widely used in a variety of computer vision and computer graphics applications, such as image denoising, edge-aware smoothing, detail enhancement, dehazing, and aggregation for stereo matching.Image filtering algorithms can be classified into global and local methods. Global methods take the operation as an energy minimization problem which demand heavy computation and are not applicable for real-time applications. The local methods are less complex, but the filtering accuracy depends on proper selection of window size. Current algorithms that use full image for filtering alleviate this shortage but suffer from attenuation of weights in rich structure or noisy regions. This project aims at overcoming the limitations of present edge-preserving smoothing methods to enhance its applicability. The overall goal of the two-year plan is to develop high-performance image filtering algorithms for the stereo matching and the dehazing problems. The annual goals of the program are as follows:For the first year, we will focus on extracting structural features of the image by using the unsupervised deep learning technology, such as the Autoencoder. At the same time, we will analyze and compare the characteristics of mathematical model of the current edge-preserving filtering algorithms to develop an effective filter. The filter should be able to utilize the structural features to enhance the smoothing effect on the non-edge regions, and reduce the effect near the edges. The algorithms will use the whole image for filtering. In addition, while ensuring high-precision filtering, efficient algorithms will be derived to reduce the computational complexity.In the last year, we will apply the techniques developed in the first year to the stereo matching and the dehazing problems. To enhance their applicability, the algorithms will be implemented using graphical processing unit (GPU) in real-time. The stereo matching procedure is used to generate a disparity map, which contains four steps: (1) computation of the primary matching cost, (2) aggregation of the cost or cost-volume filtering, (3) disparity selection via winner-take-all, and (4) post-processing for disparity refinement. The edge-preserving smoothing operator will be used in the second stage. Owing to the similarities of the smoothing requirement, a high-precision real-time image dehazing system will also be developed.

Project ID:PB10708-1942
External Project ID:MOST107-2221-E182-078