Research On Image Super Resolution Algorithm Based On Convolutional Neural Network

In the field of signal and information processing,image super-resolution reconstruction technology has been a hot research topic for many domestic and foreign scholars.When acquiring images through traditional imaging systems,it is limited by the accuracy and technical level of the hardware facilities itself,as well as the external environment and imaging.Under the influence of various factors such as conditions,the obtained image will lose the original scene information to some extent.The quality of the image has a direct relationship with the spatial resolution.The emergence of super-resolution reconstruction technology opens up a most effective solution to the problem of low image resolution from the perspective of improving the spatial resolution of the image.The basic idea of the super-resolution reconstruction algorithm is to use the signal processing method to restore the high-frequency information outside the cutoff frequency of the imaging system that is limited by factors such as imaging conditions during image acquisition,thereby improving the original image resolution.With the purpose of quality.It has the advantages of low cost and high universality.And its research results have been widely used and embodied in actual social life,such as industrial robots,security monitoring,multimedia communication technology,aerospace,mapping,remote sensing telemetry,medical aided diagnosis and many other fields.Crucial role.Common image super-resolution reconstruction techniques include reconstruction-based and learning-based methods.The essence of the former is to use the high-frequency information in the original low-resolution image and the complementary information between the multi-frame images,and combine the prior knowledge of the image to solve the inverse process of the image degradation process,which is “lost” in the imaging process.High frequency information is restored and reconstructed.However,the rebuilding ability is limited when the amplification factor increases.The latter is to learn the high-frequency information of the image and the mapping relationship between the high-and low-resolution images by training the sample image to guide the high-resolution reconstruction process of thelow-resolution image.In this paper,by analyzing the advantages and disadvantages of the existing methods for deficiencies which,taking into account the effects of reconstruction from the computing speed and angle,an improved algorithm.The nonlinear mapping convolutional layer and concat layer are added to the three-layer structure of the SRCNN(Super-Resolution Convolutional Neural Network)network model to increase the depth of the network and the number of convolution kernels,further enhance the super-resolution reconstruction effect of the image,and discard the bicubic interpolation.The pre-processing step achieves the purpose of upsampling by adding a deconvolution layer at the end of the network structure,and directly acts as an input to the convolutional neural network for images with lower resolution.For the initial image with higher resolution,in order to more intuitively compare the effects before and after reconstruction,the pooling layer is first downsampled as a preprocessing step,and then processed by a reconstruction algorithm based on convolutional neural network to output a high resolution restored image.and the better reconstruction effect is obtained through the test of structural similarity and peak signal-to-noise ratio.