Histogram of visual words based on locally adaptive regression kernels descriptors for image feature extraction

Image feature extraction is one of the most important problems for image recognition system. We tackle this by combing the locally adaptive regression kernel descriptors (LARK), bag-of-visual-words and sparse representation. Specifically, this paper makes two main contributions: (1) we introduce a novel method called histogram of visual words based on locally adaptive regression kernels descriptors (HWLD) for image feature extraction. LARK is used to describe the image local information and build the visual vocabulary. Each pixel of an image is assigned to the visual words and gets the corresponding weights. Image feature vector is obtained by subdividing the image and computing the accumulative weight histograms of visual words in these sub-blocks. (2) The K nearest neighbor based sparse representation (KNN-SR) is presented for assigning the visual words. Compared with nearest neighbors based method, KNN-SR has stronger discriminant power to identify different patches in the image. Experimental results on the AR face image set, the CMU-PIE face image set, the ETH80 object image set and the Nister image set demonstrate that our method is more effective than some state-of-the-art feature extraction methods.