Motion Blurred Image Restoration Network Based on Insulator Defect Detection

Motion blur generated during aerial photography by UAVs can lead to distortion of the target portion of the captured insulator image, which is not conducive to the subsequent positioning and defect detection of insulators. The motion blur caused by the relative motion between the aerial photography UAV and insulators is usually directional and non-uniform, and is optimized based on the Transformer model in terms of structure and objective function. The self-attention in the Transformer model is improved and combined with convolutional neural network to achieve efficient blurred insulator image restoration without relying on a large amount of training data. Contrast learning is introduced in the objective function part to fully utilize the correlation information between sharp and blurred insulator images. The motion blurred insulator image dataset is constructed for image restoration and defect detection experiments, and the results show that the proposed method outperforms the algorithms such as DeblurGAN-v2 and MIMO-UNet in terms of PSNR and SSIM, and the image quality is greatly improved.