Region-Based Near-Duplicate Image Retrieval with Geometric Consistency

State-of-the-art near-duplicate image retrieval systems take the image as a whole by the bag-of-words (BOW) representation. Feature quantization on large image database always reduces the discriminative power of image features, and the global BOW feature neglects the geometric relationships among local features. We propose in this paper a region-based image retrieval method. Image similarity is determined by the consistent MSER regions of two images, while the matched MSER regions is identified by similar local features (e.g. SIFT), where both the feature vectors and the geometric consistency among the feature locations are considered to ensure stable matching. Moreover, all the operations are realized on the original features but not the quantized ones to avoid the loss of the discriminative power. Experimental results confirm the superiority of the proposed method, especially the retrieval precision at top ranked images.