Fast alternative to gradient computation in the DCT block domain for local binary patterns computation

Computationally efficient implementation of different image processing and analysis algorithms is very important when implementing such tasks on mobile devices. Such a case can be the visual surveillance of the critical infrastructures (like hydro-dams), where images can be acquired on-site by an operator using a mobile device and an objective result of their analysis may be desired. Texture features (edge density, local binary patterns) derived from local gradient information can be efficient to assess the degradation of the concrete hydro-dam walls (in terms of crack characterization and surface degradation), and their numerically efficient implementation in the compressed image data domain allow to be implemented on any device. We propose here an efficient solution to compute local binary patterns in the compressed DCT block domain, which is the basis of JPEG image encoding, via local omni-directional gradients. Results are illustrated on an image acquired from the downstream face of a concrete hydro-dam, compared to the decompressed image implementation, and discussed in terms of numerical efficiency.