A novel shape descriptor based on salient keypoints detection for binary image matching and retrieval

We introduce a shape descriptor that extracts keypoints from binary images and automatically detects the salient ones among them. The proposed descriptor operates as follows: First, the contours of the image are detected and an image transformation is used to generate background information. Next, pixels of the transformed image that have specific characteristics in their local areas are used to extract keypoints. Afterwards, the most salient keypoints are automatically detected by filtering out redundant and sensitive ones. Finally, a feature vector is calculated for each keypoint by using the distribution of contour points in its local area. The proposed descriptor is evaluated using public datasets of silhouette images, handwritten math expressions, hand-drawn diagram sketches, and noisy scanned logos. Experimental results show that the proposed descriptor compares strongly against state of the art methods, and that it is reliable when applied on challenging images such as fluctuated handwriting and noisy scanned images. Furthermore, we integrate our descriptor in a content-based document image retrieval system using sketch queries as a step for query and candidate occurrence matching, and we show that it leads to a significant boost in retrieval performances.