The digital construction of cultural heritage promotes communication and sharing of digital cultural resources across time and space. Digital storage serves as the foundation for the digital construction of cultural artifacts. In the digital storage of Dunhuang murals, image stitching plays a critical role in restoring the complete image of the cave murals. Traditional image stitching methods are constrained by the detection accuracy of feature points and are not fit for stitching low-texture murals. Despite deep learning-based image stitching methods, parallax misalignment and ghosting are still prevalent issues. For this reason, we perform the first Dunhuang mural stitching based on deep learning in this paper. This is in response to the need for digitizing and storing Dunhuang murals. Two mural stitching datasets are constructed, and we design a progressive regression image alignment network and a feature differential reconstruction soft-coded seam stitching network. We also introduce a soft-coded seam quality evaluation method. The algorithm presented in this paper achieves state-of-the-art alignment and stitching performance in the mural stitching task through unsupervised learning with a smaller number of model parameters, which provides technical support for the digitization and preservation of Dunhuang murals. The codes and models will be available at https://github.com/MmelodYy/DunHuangStitch .
Image rectangling research aims to solve the problem of irregular boundaries in stitched images, which helps to enhance the visual satisfaction of electronic consumers with the imaging of electronic products. To do so, it learns the mesh warping rules of the image, allowing for the generation of a rectangular image with content fidelity. Existing image rectangling solutions either require two stages of warping processing, or one stage but multiple levels of warping processing to achieve the rectangling effect, both of which are inefficient at rectangling and are prone to cumulative errors and distortions of images. Based on an optimization exploration of a single-stage and multi-level warping approach, we construct an image rectangling knowledge distillation task. An image rectangling network with single-stage and single-level warping is proposed after analyzing the distillation task construction impact factors. The proposed method, combined with our multi-teacher guided instance-level knowledge distillation strategy, utilizes only 2.53% of the parameters employed by the state-of-the-art method. It achieves improvements of 4.68% in SSIM, 1.94% in PSNR, and 1.01% in NIQE metrics, while also achieving a speedup of 81.38%. The codes and models will be available at https://github.com/MmelodYy/Distill_Rectangling.
In the problem of remote sensing target detection, due to the large differences in size and shape of the target, uneven distribution of the target, and strong background interference in remote sensing images, it is difficult to identify and loca te the detection algorithm. To solve this problem, this paper starts from YOLOX. Firstly, self-calibration convolution is used to replace the deep convolution module in CSPLayer in YOLOX to enhance the context-aware ability of the model and expand the receptive field of the model. Then, in order to solve the interference of local information caused by too much consideration of the surrounding context information in the self-calibration convolution, the normalized attention module is introduced at the end of the self-calibration convolution to enhance the attention to local and non-obvious information. Finally, considering the defects of IOU in boundary box regression, GIOU loss is introduced to replace the boundary box regression. Experiments on RSOD remote sensing dat a set show that the improved algorithm improves the detection performance of 7.87 % mAP compared with YOLOX on this data set, and improves 22.03 % AP on non-obvious overpass images. In addition, compared with the current advanced detection algorithm, the improved algorithm has advanced detection performance.