Robust Differential Circle Patterns based on fuzzy membership-pooling: A novel local image descriptor

This paper presents a novel image descriptor, which is robust to a variety of photometric and geometric image transformations. Specifically, the Robust Differential Circle Patterns (RDCP) are proposed to encode the continuous intensity changes along the circular-shaped structures around each pixel. Compared to the pixel-wise feature computing schemes, RDCP is capable of describing relatively large local structures in the image. While in the descriptor constructing stage, the proposed fuzzy membership-pooling algorithm can not only capture the local structure of the interest region but also achieve rotation invariance inherently. Experimental results on three popular datasets (Oxford dataset, Patch dataset, and Ukbench dataset) demonstrate the superiority of proposed method over the state-of-the-art algorithms under various image transformations such as rotation and scaling changes, viewpoint changes, image blurring, JPEG compression, illumination changes, and image noise.