ECCA:An Efficient Channel-and-Coordinate Attention Module for Image Stitching

We propose a simple and effective attention module, called Efficient Channel-and-Coordinate Attention (ECCA), which can be applied to CNN-based image stitching models. Traditional image stitching models usually use deep convolutional neural networks, but as the depth increases, the information capacity of the model also increases, leading to unstable region of interest and degraded feature quality. Moreover, for similar features in different positions of two images, the model may mistakenly regard them as the same feature, resulting in incorrect stitching results. To address these issues, we design a multi-head attention mechanism that combines channels and coordinate positions. The module can enable the network to focus on more critical information among the numerous input information and extract higher quality features. Specifically, the module consists of two attention heads: local channel attention head and coordinate attention head. The local channel attention head calculates the correlation between each channel and its adjacent k channels, and the coordinate attention head calculates the correlation between each channel and the global channel according to the feature position in the image. The two attention heads obtain attention weights respectively, and multiply them with the original features to obtain the final output features. Our ECCA module can effectively improve the performance of existing state-of-the-art deep image stitching models. On the WarpedCOCO dataset, we apply the ECCA module to VFISnet and LB-UDHN networks, respectively improving their PSNR values by 0.63 and 1.92, and SSIM values by 5.7% and 5.1%. Code is available at https://github.com/chenwinkk/ECCA-Network