CLRF: Compressed Local Retinal Features for image description

This paper presents a new methodology to extract discriminative features from images, that are robust and invariant to image blur and JPEG compression. The local patches are quantized in the polar geometric structure using Log Polar Transformation (LPT). Then two-dimensional Discrete Wavelet Transformation (DWT) is used to decompose the polar structured patch into sub-bands. Each approximation sub-band of the wavelet decomposition is converted to vectors which are considered as features namely Compressed Local Retinal Features (CLRF). The proposed approach is comparatively evaluated with the state-of-the-art image descriptors on the standard Oxford dataset. The experimental results demonstrate the robustness of the proposed descriptor against image blur and JPEG compressions.