Deep learning white balance for Bayer raw images

Color constancy is the ability of the human eye to consistently perceive the color of objects under different lighting conditions. However, images captured by CMOS image sensors are prone to color casts, which can affect applications in related fields. Existing white balance algorithms each have their respective advantages and limitations. This paper focuses on a deep learning-based white balance method that can be integrated into the ISP pipeline. The classic ResNet34 network is trained using raw images. In this study, the activation function and loss function are optimized and trained using a stochastic gradient descent optimizer. Through processes such as image preprocessing, network training, and white balance restoration, the model is trained and validated on a dataset collected with the DJI Osmo Pocket 2. Experimental results demonstrate that the model achieves excellent white-balance effects. The system exhibits strong performance and good portability.