The Fast Multi-scale Convolutional Sparse Coding Based Super-Resolution for Infrared Image

Focused on the issue that ringing effect appeared and details are missed while improving the resolution of infrared images,a fast multi-scale convolutional sparse coding based super resolution method is proposed.First of all,the input image was decomposed into smooth component and texture detail component in multi-scales,bicubic interpolation method was used to enlarge the final smooth component in order to get the smooth component of the output image.Then,by superimposing the convolutional sum of the high resolution filters of each scale and the high resolution characteristics of its corresponding scale,the high frequency texture structure of the output image was obtained.Among them,the high resolution feature mapping of each scale was obtained by the mapping function transformation of the corresponding scale which zoomed the low resolution feature maps and had the ability to guarantee the sparsity.Filters we used can be computed as linear combinations of smaller number of separable ones and the convolution iteration process was optimized.The processing results of the infrared images obtained by general data base and our laboratory show that compared with the algorithms available,the pro-posed method improves the peak signal-to-noise ratio of the image.Besides,this method not only takes the consistency of pixels in overlapped patches into consideration,achieves high resolution image restoration,effectively re-strains the ringing effect,and preserves better image edge texture,but also effectively increases the processing speed.