DRANet: Deep Learning-Based Automatic White Balancing Approach to CVCC

초록

As a core component of the image processing pipeline, an image signal processor plays a critical role in automatic white balancing (AWB) for computer vision color constancy (CVCC). With the recent advent of deep convolutional neural network (DCNN), image signal processors have experienced enormous progress in CVCC. A myriad of deep learning-based CVCC models have emerged and outperformed their statistics-based, shallow counterparts. In this context, this article presents a novel deep learning-based AWB approach: the Dual Residual Aggregated Network (DRANet). Homogeneous Dual Residual Blocks (DRBs) are a key component of the proposed DRANet, intended to design the architecture with high cardinality. The homogeneous DRBs perform transformations simultaneously and their outputs are processed in a concatenating manner. This characterizes the proposed DRANet as a simple and wide but still deep structure which contributes to advancing illumination estimation accuracy to an innovative level and addressing the overfitting and long-term dependency issues that conventional approaches have long been struggling with. The proposed DRANet offers another advantage that comes with high cardinality, that is, a reduction in the number of parameters necessary to run the architecture. The experimental results provide compelling evidence that the proposed AWB approach has achieved an innovative progress in illumination estimation accuracy, as well as validating the invariance of both illumination and image device.

키워드