基于DCT变换和零次学习的刑侦图像超分辨率

图像超分辨率在视频侦查领域有重要作用.基于卷积神经网络的超分辨率算法通常在训练时输入人工合成的低分辨率图像,学习高、低分辨率图像的映射,很难应用于视频侦查领域.真实低分辨率图像退化过程复杂未知,且大都经过压缩算法的处理,存在人工压缩痕迹,导致超分辨率图像出现假纹理.针对真实场景下的低分辨率图像提出一种基于离散余弦变换(...

Criminal Investigation Image Super-resolution Based on DCT and Zero-shot Learning

Image super-resolution (SR) plays an important role in video-based criminal investigation. SR algorithms based on convolutional neural networks are usually trained with the input of artificially synthesized low-resolution images to learn the mapping between high-resolution and low-resolution images, and thus they are difficult to be applied in the field of video-based criminal investigation. The degradation process of real low-resolution images is complex and unknown, and most of them are processed by compression algorithms, leading to false textures in SR images. Therefore, a new SR algorithm based on discrete cosine transform (DCT) and zero-shot learning is proposed for the video-based criminal investigation images in real scenarios. The algorithm takes advantage of the repetitive similarity within the images and utilizes sub-images from the input image for training. Different from the input of the previous SR network, the proposed algorithm takes the DCT coefficients of the sub-images as the input of the SR network to avoid magnifying compression artifacts of the input image and reduce false textures. The experimental results on the standard datasets and real criminal investigation images show that the proposed algorithm can reduce the false texture caused by compression artifacts.