基于自注意力机制的图像去阴影算法

ue to the occlusion of opaque objects, shadows often exist in images, which affects the accuracy of downstream visual tasks and reduces the aesthetics of the image. Therefore, it is necessary to remove shadows from images. Existing shadow removal methods are usually based on convolutional neural networks, which are limited by the locality of convolution and cannot effectively obtain global context. To address this issue, this paper proposes an image shadow removal algorithm based on self-attention mechanism, which can borrow color and brightness information from distant non-shadow regions. Moreover, the proposed method restores color and brightness in the LAB color space to avoid interference caused by the coupling of color and brightness in the RGB color space. In addition, color and brightness losses are introduced to constrain the consistency of color and brightness. The proposed method outperforms other methods on the ISTD+ dataset, demonstrating its effectiveness.