![<?echo $_SERVER['SERVER_NAME'];?>](/template/twentyseventeen/skin/images/header.jpg)
NVIDIA is committed to promoting the development of artificial intelligence through deep learning. Start-ups in various industries can use NVIDIA's powerful AI computing platform to explore the infinite possibilities brought by artificial intelligence.
Since the first season of the "NVIDIA Startup GPU Application Contest" was officially launched in May this year, it has received active participation from member companies of the "NVIDIA Startup Acceleration Program". After NVIDIA expert review, we selected several representative GPU applications as the winners of the awards, and will continue to share their GPU innovative application stories.
High-definition images provide important details for professionals in environments such as traffic video forensics and competitive scene detail restoration. However, in practical applications, due to production technology and cost considerations, high resolution is not used in many occasions , The super-resolution camera collects image signals, so it is very important to restore low-resolution pictures to high-resolution pictures.
Image Super Resolution (SR) is a technology that converts low resolution (LR) images to high resolution (High Resolution, HR) through a certain algorithm. In the past, traditional image super-resolution technology was not satisfactory in terms of image restoration, and required a lot of manpower for comparison, which was very inefficient. The large-scale application of GPU and deep learning has brought breakthrough development to the entire industry. At present, the use of GPU to accelerate image super-resolution reconstruction has become a mature technical solution in the industry.
As a member of the NVIDIA Startup Acceleration Program, Beijing Feisou Technology Co., Ltd. (hereinafter referred to as "Feisou Technology") has a number of world-leading research results in the field of computer vision, especially in the fields of face recognition, target tracking, and target detection. By using NVIDIA Tesla P100 to conduct large-scale deep learning training on images, the company has achieved rapid reconstruction of low-quality images in video surveillance images, and the reconstructed high-definition images have good human eye sensory effects, far exceeding traditional ultra Resolution method.
The traditional image super-resolution reconstruction effect is not satisfactory
GPU brings industry innovation
In the field of image super-resolution in the early days, traditional methods were used to infer only from the surface features of the image, thereby expanding the size and resolution of the image. In academia and evaluation, PSNR (Peak Singnal-to-Noise Ratil) is generally used. The larger the value, the better). The traditional method has a significant effect on improving the PSNR value, but the image after the real reconstruction is very different from the image that the human eye sees and is expected to be restored, which makes people feel "unlike" from the senses.
Based on the above, researchers in related fields have tried to use neural networks for image super-resolution reconstruction. This has once again improved the results of PSNR, but its reconstruction speed is slow, and for complex scenes or pictures with multiple hidden features, After reconstruction, it will still feel different from the original object.
At the same time, discovering abnormalities through monitoring is a very labor-intensive task, and after an abnormality is discovered, it will also consume more energy to enlarge the picture and to understand and analyze the abnormal part more. The method not only does not improve work efficiency, but also delays time due to poor reconstruction results.
Now, with the large-scale application of deep learning and GPU, larger-scale training can be carried out on the problem of super-resolution. The deep learning model obtained not only has a higher PSNR result, but also the human eye after reconstruction. The sensory effect is also good, and it can be close to real life objects.
At present, the use of GPU to accelerate image super-resolution reconstruction has become a mature technical solution in the industry.
GPU deep learning highlights the advantages of image reconstruction
In the super-resolution reconstruction process for unclear images in the surveillance video, the surveillance personnel first find the problem in the video, and then select the frame that requires high-definition pictures for the pictures obtained by cutting frames of the video sequence, and then The picture is transferred to the deep learning model, and the deep learning model performs feature extraction on the GPU, and then performs large-scale high-resolution image reconstruction based on the extracted features. After the high-resolution picture is obtained, the required information is obtained by manually observing the high-definition picture .
Because the convolutional neural network CNN in deep learning has good feature extraction characteristics, the research team of Feisou Technology applies CNN to feature extraction in image super-resolution, and then uses methods such as transposed convolution and random channels to enlarge the image size.
Specifically, first unify the size of the picture to be enlarged, then input the picture into the CNN model, and use the random channel sorting method to enlarge the picture size. It is precisely because of the good feature extraction ability of CNN that the image has good visual sensory effects and PSNR value after reconstruction.
CNN performs a super-division diagram. Input a low-quality image with a size of fxf, and then perform a convolution operation with an output size of fxf in n layers. At the same time, the size of the convolution kernel is fixed, and the output of the last convolution layer will be random channel sorting. Characteristic map.
The schematic diagram of the enlarged picture after random channel rearrangement. Each feature map corresponds to a channel. The position of each feature value on the output image is rearranged according to the order of the channel and the size of the convolution kernel (the volume in the figure above) The size of the product core is 2x2)
New image super-resolution reconstruction technology recognized
With its outstanding performance in the field of super-resolution, the research team of Feisou Technology won excellent results in the 2018 CVPR (Computer Vision and Pattern Recognition) workshop NTIRE (New Trends in Image Restoration and Enhancement) competition.
Feisou Technology's solutions are also effective in many applications. In the example of super-resolution reconstruction of video pictures in the gaming industry, Figure a and Figure b (shown below) are the comparison of pictures before and after reconstruction. It can be seen that the algorithm model can retain a lot of detailed information after the picture is enlarged, such as the numbers on the poker cards and chips, and still has a better visual sense after enlargement.
Picture a (left) is a small low-resolution picture, picture b (right) is a large high-resolution picture
It can be said that with the strong support of NVIDIA Tesla P100 series GPUs, with the rapid development of deep learning and its application in the field of super-resolution, training larger models with more data to achieve image super-resolution reconstruction has become reality. Not only the training speed is fast, but the running speed of the model after the training is several times to several tens of times faster than the CPU, the PSNR index has also been greatly improved, and it is closer to the lines of objects in normal life.
Feisou Technology's use of Tesla P100 to achieve super-resolution image reconstruction is only the tip of the iceberg of applications in this field. In the future, there will be larger scale, more representative data and better hardware to further improve image super-resolution in practical applications Ability and performance in
Dongguan Guancheng Precision Plastic Manufacturing Co., Ltd. , https://www.dpowergo.com