Corner Detecting And Image Matching Algorithms Of Low Quality Images

Corners are essential image features with meaningful and important information, and thus have been widely used in image processing and computer vision. Corner feature detection and image matching are particularly important to the video image analysis, understanding, and intelligent interpretation in the intelligent monitoring and detecting system. Image matching can be divided into three kinds which are the gray based image matching, the transform domain based image matching and the feature based image matching. The feature based image matching is a hot topic because it has low complexial computation and is robust to noisy, and also can meet the complex geometric transformation.Many video images are polluted by noise and non-uniform illumination, and the video image quality is poor because of the complex working environment. The performance of feature extraction and image matching is low on this case. At the same time, more and more systems also need solve the real-time demand. Although there has been a large number of feature detection and matching algorithms, it has a lot of problems of the performance and time efficiency for low quality images. This paper proposed some corner detection algorithms and corner based image matching for low quality video images. The algorithms focused on improving the feature detection and matching performance, and reducing the time consuming.We present an adaptive corner detector based on multi-scale chord-angle sharpness accumulation (AMCSA). It can effectively solve the anti-noise problem. Firstly, the algorithm improvs the canny edge detection and edge location accuracy. It uses low edge response inhibition to speed up the edge contour extraction efficiency. Then, the corner support domain is divided into three aspects which are as three scales. For each scale, the mean chord-angle sharpness is calculated, and the three chord-angle sharpnesses are accumulated as a multi-scale corner response function. The multi-scale chord-angle sharpness accumulation can effectively enlarge the response difference between the true corners and noisy-points, and can greatly improve the anti-noise ability. The algorithm does not smooth the contour during the calculation of the corner response function, and can keep high location accuracy. Finally, an adaptive threshold of each edge is calculated separately according to its own characteristics to improve the adaptability of the algorithm. The experimental results show that, the proposed algorithm can effectively suppress noise, and improve the corners location precision, and also can solve the fuzzy problem to some extent. The adaptive threshold makes the algorithm have better adaptability.However, The AMCSA algorithm excessively relies on the extracted edges, and is also time-consuming. It performs poorly on non-uniform illumination and fuzzy images. In this paper, we proposes a fast corner detector for non-illumination and fuzzy image based on dual-threshold (FIFD) to improve the performance of low quality video images. Firstly, we presents an inner mask which uses only four pixels to determine the flat regions and corner regions of an image, which can get rid of the unnecessary computation on flat regions to reduce computing cost. Secondly, we separate the corner regions into background and foreground and compute the separate corner-threshold to settle non-uniform illumination. The thresholds are computed adaptively according to itself characteristics and through training the video images to speed up the double threshold calculation to effectively solve the non-uniform illumination. Thirdly, we propose a fast corner detection algorithm to compute nucleus continuous contributive segment based on the corner state. Some differences pixels are concluded in the continuous contributive segment to deal with the anti-noise and fuzzy problem. Finally, we propose two effective methods to remove the false corners. The results on Caltech image, natural scenery and mine video images show that this algorithm takes the least time consuming and performs well and reliably in the location accuracy and the detection performance aspects on the non-uniform illumination and fuzzy images.Based on the FIFD corner detection, we propose a fast image matching algorithm based on the corner state and global consistency. Firstly, the algorithm maps the pixels of the neighborhood into gradient space, and computes the weighted mean distance of the pixels in the gradient space to straight line. And the line which has the maximum weighted mean distance is considered as the main direction of the corner. The main direction is unique, and is invariant to the scale and rotation, and also is computed very simple. On this basis, the binary texture features and statistical histogram are combined as a descriptor vector to strengthen the distinguishability for the low quality images. Then we adopt a coarse to fine matching strategy to improve the matching speed. Finally, according to the space position relations between the matching points, the matching points which meet the affine transformation matrix are remained. The method of minimum error fluctuating range clustering is used to quickly remove the error matches. The experiment results show that, the matching algorithm is invariant to image rotation, illumination changes, affine transformation, noise and fuzzy and so on. It performs well in the performance and efficiency. And also can be effectively applied to the image mascio and object detection of the videos in complex conditions.