Research On Convolutional Neural Network-based Image Interpolation Filtering Method And Application

With the great achievement of the internet and the fifth-generation communication technology,ultra-high-definition images and videos are becoming more and more popular.Limited by factors such as acquisition equipment and environment,it is usually hard to directly obtain super-high resolution images or videos,thereby requiring post-processing for super-resolution.At the same time,compared with low-resolution images and videos,the amount of data contained in ultra-high-definition images and video content is very huge,therefore,to further improve the compression efficiency of high resolution videos and images is another urgent task.Interpolation filtering is one of the important research directions in the field of signal processing,and plays a key role in image super-resolution tasks and video compression tasks.Image super-resolution is the process of interpolating low-resolution images to obtain high-resolution images.The performance of the interpolation filter determines the quality of the high-resolution images.In the task of video compression,fractional samples are obtained by interpolating and filtering the integer pixel samples to realize higher precision fractional motion compensation.As the demand of high quality images and videos is becoming higher and higher,traditional interpolation filters are hard to adapt well to images and videos with various content.Recently,convolutional neural networks have been widely used in computer vision tasks,especially in image segmentation,super-resolution and other fields have made great progress,which provide a new direction for the research of adaptive interpolation filtering.Therefore,we introduces the convolutional neural network into the design of the adaptive interpolation filter,proposes a new image super-resolution model,and applies it to the new generation video coding standard H.266/VVC(Versatile Video Coding)coding platform VTM7.0.The main content is as follows:(1)Inspired by the concept of static operating point in triode amplifier circuit,we propose a new type of activation function.The zero point in the original ReLU function is taken as an operational and learnable point during the neural network training process.It can effectively alleviate the neuron necrosis problem caused by improper parameter initialization or excessive learning rate of the original ReLU function.By applying it to the existing image super-resolution model,obvious gains can be achieved.(2)Many existing image super-resolution networks based on the feedforward structure adopt a single up-sampling structure.They only use the up-sampling module at the end of the network to up-sample the learned low-resolution features.However,such approaches does not fully address the mutual dependencies of low-and high-resolution images.Therefore,we combine the ideas of residual network and projection,and proposes a residual projection module(Residual Projection Block,RPB).Based on this module,this paper proposes a new image super-resolution model(Residual Projection Network,RPNet).Experimental results show that the model can achieve better results on the commonly used data sets.(3)By taking the video coding distortion into account,the proposed RPNet is retrained and applied in the coding platform VTM7.0 of the new generation video coding standard H.266/VVC(Versatile Video Coding),for half-precision fractional sample interpolation.Experimental results demonstrate that this model can improve the accuracy of fractional motion estimation,and thereby improving the video coding performance.