Super Resolution Reconstruction Of Traditional Surveillance Video Image Based On GPU

With the development of information technology,surveillance video images have become one of the important channels for people to obtain information about the surrounding environment,and have played a very important role in people's lives.However,the surveillance's equipment is too obsolete,and the video has been encoded and compressed,often makes the captured surveillance video images have low resolution and difficult to identify details,which has caused many problems to our lives and production.Researchers use image super-resolution reconstruction technology to reconstruct these low-resolution surveillance video images to obtain higher-resolution surveillance video images with higher resolution and better image quality.Although there have been many research results,the final reconstructed image quality is still not very satisfactory,and there is still room for improvement.At the same time,because the super-resolution reconstruction algorithm has a large number of complex calculations,it is difficult for general embedded processing platforms to meet its requirements,resulting in a very slow operation speed,which makes the super-resolution reconstruction of traditional surveillance video images slow in the application of embedded engineering environment.Against to the above problems,this paper uses the convolutional neural network method to perform super-resolution reconstruction of traditional surveillance video.It is mainly for super-resolution reconstruction of objects and related motion information that may exist in traditional surveillance video images.At the same time,considering the embedded operating environment of traditional surveillance video images,the super-resolution reconstruction neural network was transplanted to the embedded GPU platform TX2.The specific research work in this paper mainly includes:(1)Based on the analysis of the traditional surveillance video image degradation model,a method based on convolutional neural network for extracting image spatial information and motion information is applied to the traditional surveillance video image super-resolution reconstruction problem,which leads that super-resolution reconstruction convolutional neural network is designed.The neural network extracts and fuses spatial information and motion information in multi-frame low-resolution video images to reconstruct images with higher resolution and better image quality.Through experimental comparison,it is proved that the convolutional neural network proposed in this paper has better super-resolution reconstruction effect on traditional surveillance video images,and further improves the quality of traditional surveillance video images.(2)Considering the traditional surveillance video embedded engineering environment and the requirements of the super-resolution reconstruction convolutional neural network for hardware computing capabilities,the super-resolution reconstruction convolutional neural network algorithm proposed in this paper is transplanted to the embedded GPU platform NVIDIA Jetson TX2.(3)Since the TX2 platform has smaller memory and weaker computing power,the original neural network has a longer running time,which needs to be further optimized to improve its adaptability on the TX2 platform.On the one hand,this paper optimizes the structure and operation method of the convolutional neural network to reduce its memory space and calculation amount.On the other hand,under the premise of ensuring that the quality loss of the super-resolution reconstruction image is acceptable,this paper combines the TensorRT neural network model reasoning optimization to significantly improve the super-resolution reconstruction speed of the convolutional neural network.Therefore,the convolutional neural network has more engineering application value.