Guiding the Illumination Estimation Using the Attention Mechanism

Abstract

Deep learning methods have achieved a large step forward in many computer vision applications. With mechanisms such as attention, deep models can now guide themselves to focus on parts of an image that are more significant for a given task. In computational color constancy, the most important step is to estimate the illumination vector as accurately as possible. Since illumination estimation algorithms can be sensitive to noise, such as ambiguous regions in the image, the ability to have a mechanism to look for specific regions in an image could be helpful. In this paper, a convolutional neural network with an attention mechanism is proposed. The attention mechanism helps the network to focus on regions that contain more content and to avoid regions where ambiguous estimations may occur. In the experimental results, it is shown that the attention mechanism does help the network to obtain more accurate estimations and puts the focus of the network on the regions in an image where gradients are high. The network with the attention mechanism achieves up to 10% increase in accuracy compared to the same network architecture without the attention mechanism.