Research On Super-Resolution Of Image Based On Extreme Learning Machine

Image is an important information that we perceive and understand the outside world.However,the image information we obtained in many practical situations is the image with low resolution.Because the high resolution image can provide rich details and sensitive color conversion,it is necessary to develop a low cost and easy to implement image processing technology to improve the resolution of the image.Image super resolution is a technique that transforms a low-resolution image of one or more small and blurred images into a high-resolution image with large,clear and more rich details of the image.Image super resolution technology has been widely applied in many fields,such as medical devices,intelligent transportation,text identification,security monitoring and so on.From machine learning point of view,image super resolution is essentially a mapping relationship from low resolution image to high resolution image.Extreme learning machine is a nonlinear mapping tool with excellent performance.It has been widely applied in machine learning,pattern recognition and other fields.Based on the extreme learning machine,we have studied the technology of image super resolution.The main contents are as follows:First,in order to solve the problem of lack of generalization performance in the existence of over-fitting of the fixed structure extreme learning machine,an image super-resolution algorithm based on L1 regularization based extreme learning machine is proposed.The algorithm uses the least angle regression algorithm to solve the problem of the L1 regularization extreme learning machine and in the final stage of image reconstruction,the global image information is restored by iterative back projection.It is proved by experiments that the method can achieve good effect in the process of image super-resolution.Secondly,to deal with the drawback of the image super-resolution algorithm based on neighborhood embedding,which only takes into account the linear combination of adjacent blocks without considering the nonlinear combination of image blocks,a nonlinear neighborhood embedded image super resolution method based on extreme learning machine is proposed.The method uses extreme learning machine to learn the nonlinear relationship between image blocks and applies it to image block matching.The experimental results show that the proposed algorithm has improved both in visual effect and in the numerical representation of contrast method.Finally,to overcome the shortcomings that lack of network depth and can not achieve more complex nonlinear mapping of the image super-resolution method based on the extreme learning machine,an image super-resolution algorithm based on depth representation extreme learning machine is proposed.In this method,multiple extreme learning machines are connected in series,then a deep extreme learning machine algorithm is obtained,which is applied to image super resolution operation.Therefore,the experimental results are significantly improved compared with other methods.