Decoupling Semantic Context and Color Correlation with Multi-class Cross Branch Regularization

This paper presents a novel design methodology for architecting a light-weight and faster DNN architecture for vision applications. The effectiveness of the architecture is demonstrated on Color-Constancy (Auto White Balance) use case, an inherent block in camera and imaging pipelines. Specifically, we present a multi-branch architecture that disassembles the contextual featuresand color properties from an image, and later combines them to predict a global property (e.g. Global Illumination). We also propose an implicit regularization technique by designing a cross-branch regularization block that enables the network to retain high generalization accuracy. With a conservative use of best computational operators, the proposed architecture achieves state-of-the-art accuracy with 30X less model parameters and 70X faster inference time for color constancy. It is also demonstrated that the proposed architecture achieves similar efficiency in case of other low-level image restoration problems such as Low-Light enhancement.