Decoupling Semantic Context and Color Correlation with multi-class cross branch regularization

TL;DR

This paper introduces a lightweight multi-branch neural network architecture that effectively separates and combines semantic context and color information, achieving state-of-the-art accuracy with significantly reduced parameters and faster inference for color constancy and other vision tasks.

Contribution

The paper proposes a novel multi-branch architecture with cross-branch regularization for efficient and accurate vision models, especially in color constancy.

Findings

Achieves 30X fewer parameters than existing models.

Provides 70X faster inference time.

Maintains high accuracy across multiple vision tasks.

Abstract

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 use case an inherent block in camera and imaging pipelines. Specifically, we present a multi-branch architecture that disassembles the contextual features and 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 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 lesser model parameters and 70X faster inference time for color constancy. It is also shown that the proposed architecture is generic and achieves similar efficiency in other vision applications such as Low-Light photography.