One-net: Convolutional color constancy simplified

Abstract

• Proposed model achieves state-of-the-art results on tested datasets. • Proposed neural network is simple and suitable for real-time applications. • Simple image augmentation results in improved cross-camera illumination estimation. • Architecture with small kernal outperforms larger kernels in speed and accuracy. Images have an ever-increasing presence in our daily lives. This increases the need for accurate and efficient image processing . One of the first processing steps in modern cameras is image white-balancing, the process of making the image invariant to the illumination of the scene. This can be achieved by estimating the illumination of the scene, which is used to chromatically adapt the image. Many existing state-of-the-art approaches use pre-trained models as feature extractors. These models are pre-trained on ImageNet and usually have several million parameters. In this paper, we introduce a simple convolutional neural network without pre-trained layers, that achieves state-of-the-art results. The model contains five convolutional layers , and all of them have a small kernel of size (1,1). Experiments with different model complexities and different kernel sizes have shown that high-level semantic information obtained using larger kernels is not required to achieve state-of-the-art results. Cross camera experiments were also performed and they showed that simple image pre-processing can significantly decrease the effect of camera-sensor on the method. The proposed method has less than 22 000 parameters and achieves state-of-the-art results. The model was tested on three different datasets: the Cube+ dataset, the NUS-8 dataset, and the Intel-TAU dataset.