Alternating direction method based decoding for object recognition

Image-based object recognition suffers from the corruptions caused by lighting, contrast, occlusion, and other image noises, which are treated as errors usually. The errors are often with high dimensions. Sometimes, they are modeled as additive noises composed of the sparse error (e.g., occlusion) and the Gaussian noise (e.g., cluttered background). This paper proposes an Alternating Direction Method (ADM) based algorithm called ADM-decoding for the error correcting (decoding) problem when the Gaussian noise exists. Our algorithm is with low complexity and decomposed into two parts. One is to solve an optimization problem by soft threshold function, which has been widely used in sparse recovery. Another concerns some simple operations of matrices to solve an optimization problem else. Simulations are given to show that the ADM-decoding is more suitable than some existing algorithms for reconstructing the object signal from highly corrupted measurements in high-dimensional cases. Also, it is availably applied to some certain object recognition problems, such as background subtraction and robust face recognition.