Research On The Binary Descriptors Of Image And Object Recognition Algorithm

Image feature extraction and anlysis is an important topic in computer vision. It has been widely used in object recognition, image classification and matching, and face recognition etc. Floating-point feature occupies large memory and high computaional cost, so it can not meet the increasing large-scale image data handling requirements. While binary feature descriptor is simple and efficient, so it has wide applications in image processing and computer vision. However, binary descriptor is sensitive to noise and the quantization is too coarse, so it loses much imformation and lacks enough discriminative ability. The binary descriptor has performance gap compared with the floating-point descriptor. We have analyzed the current binary feature descriptors and have improved the performance. In this dissertation, we propose several novel binary feature descriptors based on the region invariance for object recognition applications.Firstly, we have analyzed the distribution of the intensity differences between the eight adjacent neighbors and the central pixels of all the local affine covariant regions and found that it nearly follows the standard normal distribution. Then we propose the local adaptive derivative quantized binary pattern. In order to incorporate the spatial information and improve the robustmess to lighing changes, we partition the local affine covariant regions into several sub-regions according to the intensity orders instead of the square-based partition. Since the distribution of the intensity differences between the eight neighbors and the central pixels of all the local affine covariant regions is nearly a standard normal distribution, we use the standard deviation as the threshold and make adaptive quantization levels. To improve the robustness to rotation, rotaion invariant sampling is utilized and the differces are computed. Then the differences are sorted and mapped into binary codes with a Hash funtion. We then transform these binary codes into decimals and count the frequencies of the decimals to form the histogram. Furthermore, Multi-scale image descriptor is utilized to improve the distinctiveness. We change the size of the local affine covariant regions and concatenate the histograms together and form the multi-scale descriptor to represent the feature the local affine covariant regions.Secondly, we have seen that the local descriptors describe the feature of the stable regions under different deformations, and they lose the global structure information. Since the intensity differences between the eight neighbors and the central pixels of an image also follow the standard normal distribution, we propose the local gradient ordinal binary pattern to represent the global featue of the image. The whole image is divided into several sub-regions according to the intensities. For each pixel in the sub-region, we compute the gradients of 8 orientations and sort the gradients. Then the sorted gradients are quantized according to global standard deviation and mapped into binary codes. They are transformed into binary codes and histogramed with a wighting function. We change the neighbor size and get the multi-scale descriptors. The multi-scale descriptors are then pooled together to represent the global feature of the image. Compared with the local descriptor, global feature can preserve the spatial information of the image.Finally, the traditional binary features usually perform binarization on the intensity difference and thus lose some information for the quantization is too coarse and they are not suitable for object recognition. We have seen that the higher four bit planes contain the main information of an image and they are robust to noise, so we propose the rotation invariant significant bit-planes-based local binary pattern. The whole image is divided into several sub-regions according to the intensities. For each pixel in the sub-region, we sort the neighbors and extract the four higher bit planes respectively, which avoid the information lost and make the descriptor robust to rotation. Multi-scale descriptor is used by changing the size of the neighbors. We average the adjacent pixels in the larger scale and then extract the four significant bit planes respectively, which is robust to noise and captures more information and further improve the discriminative ability of the descriptor. We then concatenate the multi-scale descriptor to form the global feature of the image.Our descriptor is robust to lighting changes, scale variation and rotation. We propose the similarity measurements based on the binary descriptors and the object recognition algorithm. We have achieved high recognition accuracy on three benchmarks and the results show that our descriptor outperforms state-of-the-art binary descriptors and SIFT.