Binary Image Feature And Its Applications

Extracting effective image feature is a key step in many computer vision prob-lems, which usually determines the success of a computer vision method. With the widely access of internet and handsets, the real-time processing of the large-scale image or video materials has become a great challenge in the area of computer vision even the whole computer science. Thus, it gives a higher level requirement for the image feature extraction. Converting the feature to binary codes or directly extract-ing binary features from image gives a good solution to the large-scale and real-time computer vision problems. This paper exactly focuses on obtaining effective binary image feature for different computer vision problems.Firstly, we propose a semi-supervised hashing algorithm to convert the exist-ing feature to binary codes. We choose the anchor-based nonlinear hash function to improve the generalization ability. Both labeled and unlabeled data are considered during the learning. Moreover, a bootstrap sequential learning method is proposed to effectively correct the errors by taking into account of all the previous learned bits holistically. Experimental results on the six benchmark datasets demonstrate that the performance of the presented method.Secondly, we propose a convolutional treelets binary feature method for fast real-time keypoint matching. In our method, we directly formulate the keypoint recog-nition as an image patch retrieval problem, which enjoys the merit of finding the matched keypoint and its pose simultaneously. A novel convolutional treelets ap-proach is proposed to effectively extract the binary features from the patches. A corresponding sub-signature-based locality sensitive hashing scheme is employed for the fast approximate nearest neighbor search in patch retrieval. Experiments on both synthetic data and real-world images have shown that our method performs better than state-of-the-art descriptor-based and classification-based approaches We also introduce randomly projected binary feature in the video copy detection. A very efficient sparse random projection method is employed to encode the image features while retaining their discrimination capability. To tackle the tough trans-formations, several effective preprocessing techniques are introduced. We present a keyframe-based copy retrieval method. Moreover, an effective scoring and localiza-tion algorithm is proposed to further refine the retrieved copies and accurately locate the video segments. The promising results in the TRECVID2011content-based copy detection task demonstrated the effectiveness of our proposed approach.