A NOVEL IMAGE RETRIEVAL APPROACH BASED ON SEMANTIC AFFINITY

Feature expression and similarity metric are two crucial steps for content-based image retrieval. In this paper, an novel image retrieval approach based on semantic affinity is proposed through exploiting discriminative features and capturing the latent data manifold. Different from the traditional practice, firstly Axiomatic Fuzzy sets (AFS) is introduced to masterly embed the semantic into original image features for forming a new semantic feature space. Then different semantic descriptions are exploited for different images according real image attribute distribution, which effectively reflect the semantic difference between images. Finally, a semantic affinity metric is defined for image matching based on the similarities of semantic descriptions. It is noteworthy that this proposed metric consider not only the similarities between two images but also their their respective neighbors, which is totally different from the traditional distance metrics model only base on pair-wise distance of images. With accurate description and reflection of semantic affinity, the proposed method can further enhance the stability and robustness of similarity relationship of images and thereby leads to improvement of the retrieval performance significantly. Extensive experiments on benchmark databases verify that it can achieve a good performance in completing the face retrieval task based on pixel-level grayscale features and the natural image retrieval task based on CNNs depth features.