An improved product quantization method applying fuzzy clustering

With the development of computer vision as well as retrieval technology, content-based image retrieval (CBIR) has been widely used. The product quantization retrieval algorithm (Product Quantization) is one kind of CBIR, which contains three steps: Cartesian integral solution of the original vector space, quantization of the lowdimensional vector space, and similarity search. The original product quantization method uses k-means clustering algorithm for quantization of low-dimensional vector space, which will be influenced by the initial point setting to reach local optimum and cause quantization error. Moreover, the features of many pictures are mediated, and the k-means algorithm for the clustering of feature vectors belongs to the hardening score, which will produce a large classification error. To address the above problems, this paper proposes to apply the improved product quantization method of fuzzy clustering (Fuzzy Product Quantization) and tests the effects of different quantization parameters on retrieval accuracy. The quantization of the low-dimensional vector space by using fuzzy clustering avoids the quantization error caused by hardening scores and makes the classification more objective. The use of weighted asymmetric distance calculation in a similar search step can further reduce the error of distance calculation and improve the accuracy of image retrieval. Experimental tests using the Oxford Building dataset show that the retrieval accuracy of the original product quantization method is 61.8% when taking the first 20 images in the retrieval results, and the retrieval accuracy of the proposed method is 64.8%, with an overall accuracy improvement of 3%, and the algorithm performance is better than the original product quantization method.