基于改进上采样技术的图像超分辨率重建

Image Super-resolution Reconstruction Based on Improved Upsampling Technology

图像超分辨率重建技术一直是计算机视觉领域的热门研究方向.为了提高重建后图像的质量,本文提出一种基于内容感知的上采样技术用于图像的重建.将稠密残差网络作为骨干网络,用基于内容感知上采样取代传统的亚像素卷积上采样技术,即在特征重建阶段,卷积核不会在整个特征图中共享参数,而是神经网络可以根据特征图的内容在每个像素处生成特定的卷积核.该算法减少了参数数量,从而加快了网络训练速度.通过多轮训练和测试,结果显示使用改进技术得到了更加清晰的重建图像,取得了良好的视觉效果.

Image super-resolution reconstruction technology has always been a hot research direction in the field of computer vision. To improve the quality of reconstructed images, this study proposes an upsampling technology based on content awareness for image reconstruction. The residual dense network is used as the backbone network, and the content awareness-based upsampling replaces the traditional sub-pixel convolution upsampling. In other words, in the stage of feature reconstruction, the convolution kernel will not share parameters in the entire feature map, but the neural network can generate a specific convolution kernel depending on the content of the feature map in each pixel. The algorithm reduces the number of parameters, thereby speeding up the network training speed. After multiple rounds of training and testing, the results show that the improved technology can yield a clearer reconstructed image and presents a great visual effect.