Color-Distorted Image Classifier Based on Combined Dilated Convolution and Deployment on Smartphones

White balance is a critical component in portable consumer electronics equipped with cameras, such as digital cameras and smartphones. To further enhance the quality and efficiency of image processing, this study introduces a convolutional neural network (CNN)-based classifier into the white balance process, enabling white balance functionality. This paper proposes an automatic white balance classification network named AWBDILNet, which builds upon existing color-distorted image classification models. AWBDILNet replaces standard convolutional layers with combined dilated convolution, marking a novel application in white balance correction. It integrates an activation function that combines ReLU and H-swish, and features a simplified network structure. Additionally, the number of neurons in each layer is optimized for the binary classification of color-distorted images. An intelligent white balance processing system is developed using the AWBDILNet classifier. In this system, normally balanced images remain unprocessed, while color-distorted images undergo correction via histogram shifting. The complete system is deployed on smartphones. Experimental results show that AWBDILNet achieves a processing rate of 46.2042 Hz, which is 33.50% higher than that of the best-performing traditional method, ShuffleNet-v1. The AWBDILNet classifier effectively reduces computational complexity while maintaining high classification accuracy. Furthermore, it enables efficient and intelligent white balance correction on smartphones.