Research On Mine Image Super-Resolution Reconstruction Algorithm Based On Feedback Residual Network

As the concept of“intelligent and safe”precision coal mining is proposed,the coal industry is being transformed from the automation of manual workers to the automation of knowledge workers.Recovering underground scenes with high definition and low delay has become one of the important issues to be solved in coal intelligent mining.Due to the special environmental factors such as dust,light,and the limited hardware conditions of video equipment,mine images generally have problems such as low resolution and poor visual effects,which seriously affects the acquisition and use of detailed information of downhole scenes.Super-resolution reconstruction technology can improve the resolution of mine images,restore the lost information,and plays an important role in the intelligent mining,unmanned inspection and other computer vision tasks in the coal mine.This paper mainly studies the super-resolution reconstruction of mine images.After studying super-resolution reconstruction algorithms based on convolutional neural network,the feedback residual network module is constructed by combining the residual network structure and the reconstruction process of different frequency components of the image.Then,this paper proposes two mine image super-resolution reconstruction models based on feedback residual network,the main work and innovation are as follows:（1）For the problem of large number of parameters existing in the super-resolution reconstruction algorithm based on the convolutional neural network and the loss of edge details when reconstructing blurred mine images with weak edge features,this paper proposes a two-channel feedback residual network super-resolution reconstruction algorithm based on edge guidance.The algorithm combines the characteristics of the mine image,uses the phase consistency method to extract the low resolution edge feature and combines it with the original low resolution image as the input of the network to direct the reconstruction of the edge details of the mine image.Local and global residual learning are used to reduce the burden of carrying information during network training so that the network can be easier to train.This network can also keep the number of parameters constant while increasing network depth through recursive learning.Edge loss and reconstruction loss is introduced to construct multi-task loss function to optimize network model parameters.Experiments show that compared with other super-resolution reconstruction algorithms based on deep convolutional neural network,this algorithm can better restore the edge features of the mine image,improve the reconstruction performance,and reduce the number of network parameters and the difficulty of training.（2）Some super-resolution reconstruction algorithm based on convolutional neural network needs to perform interpolation pre-processing before entering the low-resolution image into the network and needs to train different networks for different reconstruction factors.In order to solve these problems,this paper proposes a multi-scale feedback residual network reconstruction model based on sub-pixel convolution.This model uses the redundancy of low-frequency information to set the multi-scale preprocessing module at the front end of the network,and applies the feedback residual module to training networks of different magnifications.At the end of the network,a sub-pixel convolutional layer is introduced to replace the input image preprocessing step.At the same time,the PReLU activation function and ₁l loss function are used to increase network convergence speed and improve reconstruction performance.Experiments show that the network can reduce the mine image reconstruction time,and improve the network applicability through multi-scale reconstruction while ensuring a good reconstruction effect.