Research Of Image Retrieval Based On Spatial Relationship Among Adjacent Keypoints

Bag-of-Word (BoW) model plays an important role in content based image retrieval framework. Although this model had obtained good retrieval performance, lots of mismatches are still generated due to the quantization errors and the loss of spatial information. To alleviate the problem, binary code embedding is introduced to the BoW model. In addition to representing each local feature with a visual work, a compact binary code is also associated with the local feature. However, exiting embedding schemes is to simplify the local feature in essence, and no spatial information is taken into account. To address this problem, we propose three light-weight embedding schemes to encode spatial information, which can be summarized as follows:(1) Spatial Information Embedding Based on Content Similarity between Adjacent Keypoints (CSE)To fully exploring the content similarity and spatial distribution between any key point and its adjacent key points, we propose a content similarity embedding scheme. In particular, we extract the nearest M local features surrounding k and divide the surrounding region into8equal portions started from the dominant orientation of current local feature k. Finally, a binary code is generated by encoding occurrence of neighbors in each portion in counterclockwise.(2) Spatial Information Embedding Based on Scale Similarity between Adjacent Keypoints (SSE)To fully exploring the scale similarity and spatial distribution, we propose a scale similarity embedding scheme. For any local feature k in image, we first select M nearest neighbors. Then,4local features whose scale is closest to the scale of k are selected from the M nearest neighbors. According to the distance between k and its Mth neighbor, we make a circular region as k’s neighborhood region, which is partitioned into a set of patches. By encoding occurrence of scale-similar neighbors in patches, we can generate a compact4-bit binary code.(3) Spatial Information Embedding Based on Location Similarity between Adjacent Keypoints (LSE)To fully exploring the location similarity and spatial distribution, we propose a location similarity embedding scheme. For any local feature k in image, we extract the feature descriptor and find out its nearest N local features. And then, we extract their coordinate information and compare the spatial location information between local feature k and the N adjacent points. Once obtaining the relative position information, we can finally build two N binary code according to the X, Y coordinates.