Research Of Local Image Descriptor Based On Intensity Invariance

Image local invariant feature is composed of the pixels that intensity values change greatly, they contain extremely rich image information, with strong representation. Local invariant feature extraction and description is the research basis of many problems in the areas of image analysis and computer vision, such as digital watermarking, image processing, video retrieval, target recognition and tracking, it is an important research content. The images are mostly suffered from some generalized affine transformation, such as rotation, viewpoint, blur, scale, illumination, etc., therefore the issue that how to obtain a local invariant feature with good stability, matching and reproducibility becomes an emphasis in related fields. Generally, the methods based on local invariant feature achieve their goals through the feature extraction (including feature detection and description) and feature matching. Most of the existing descriptors use gradient statistical characteristics to detect local invariant region, they can only guarantee the invariance to linear changes in intensity, which greatly limits its application in practice, because the light conditions are different and the image intensity conversion is usually not linear. Through in-depth study of the detection and description of local invariant region, we construct new descriptors, which have the invariance to monotonic intensity transform, even have better robustness to more complex nonlinear light transform.Center-Symmetric Local Binary Pattern (CS-LBP) is a widely used descriptor for local invariant feature, it is invariant well to the linear illumination transform, and need less storage space and computing cost. CS-LBP descriptor is constructed in the single support region, but sometimes deviation may occur when we choose the size of the region, the selected size of the region is too large or too small, which leads to the constructed descriptor cannot guarantee high accuracy in feature matching. To solve this problem, combined with the idea of multiple support regions, we propose a new descriptor with intensity invariance. Firstly, we extract local invariant feature regions in different sizes, and then standardize them into multiple support regions. Next, we compute a local invariant feature descriptor for each support region. Finally, we obtain a new local invariant feature descriptor by combining these descriptors of multiple support regions. The experiments of feature matching show that the new descriptor is able to maintain good invariance to illumination transform, and also get higher matching accuracy in images with JEPG compression, blur, zoom, rotation transformation.Most of the existing descriptors’construction need to estimate the main direction of the gradient statistics, the accuracy of the main direction estimation directly affects the uniqueness and stability of the descriptor, and the region dividing methods according to the main direction use the gradient statistics, result in having no invariance to monotonic intensity transform. To solve this problem, this dissertation studies the principle of intensity orders relation, and constructs a new local feature descriptor with intensity invariance by using intensity order, which is the underlying information of images. Firstly, compute the CS-LBP values of all pixels in each support region. Then, the pixels in each support region are grouped into n bins where each bin has pixels with similar ordinal pixel intensities, and draw the histogram of the number of each CS-LBP value in each intensity order bin. Finally, combine all the histograms into one histogram, the statistical values of this histogram are ranked in the form of row vector, the vector is our new invariant feature descriptor. The feature matching results of different brightness transform images show that our descriptor has good brightness transformation invariance, and also get higher matching accuracy for images with zoom, JPEG compression transformation, all of these prove the effectiveness of our new descriptor.