News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
用机器学习方法提高图片的清晰度
Improve the clarity of pictures by machine learning
Improve the clarity of pictures by machine learning
One, background
With the increasing resolution of televisions and moving screens, people’s demand for high-definition pictures has also grown. But this also creates a certain turbulence for the user - to see the big picture in high definition means that it takes up a lot of bandwidth, one is the increase in data costs, and second, the loading speed will slow down, resulting in poor user experience.
The
Therefore, how to achieve high-definition effects through the transmission of small images without affecting the user experience is a very worthwhile discussion. In October 2016, Google published a paper about their introduction of a new technology called Rapid and Accurate Image Super-Resolution (RAISR), which uses machine learning to convert low-resolution images into high-resolution images. This technology can reach resolutions and even exceed the original image at a bandwidth of 75%. At the same time, the speed can be increased by about 10 to 100 times.
Second, demand analysis
function:
1. Low-quality web pictures are converted into high-definition pictures
2. Stop deflation of high-definition pictures on the cloud and reduce the size of pictures
Third, complete the plan
The hyperfraction model construction diagram is as follows:
Google uses machine learning to exercise the program with a pair of low-resolution, high-resolution pictures to find filters that can be selectively applied to each pixel in a low-resolution picture, which can produce details that are comparable to the original picture. There are currently two ways to exercise RAISR:
· The first is a "direct" approach where the filter learns directly in pairs of high and low resolution pictures.
The second approach requires the application of low-power upsampling on low-resolution pictures first, followed by learning filters in the combination of up-sampled pictures and high-resolution pictures.
· The "direct" approach is faster to dispose of, but the second approach takes care of non-integer-wide elements and is better off with hardware performance.
Either way, the RAISR filter is based on the edge features of the image: brightness and color gradients, level and texture areas. This is affected by direction (edge angle), strength (strength with higher sharp edges), and coherence (a measure of the quantification of edge directionality). The following is a set of RAISR filters learned from 10,000 pairs of high and low resolution pictures (low resolution pictures are upsampled). This exercise takes about 1 hour.
Note: 3x super-resolution learning, achieved 11×11 filter convergence. Filters can be learned from a variety of super-resolution elements, including local super-resolution. Note that when the angle of the edge in the figure changes, the filter angle also rotates. Similarly, as the intensity progresses, the sharpness of the filter also improves; as the viscosity progresses, the filter's anisotropy also improves.
From left to right, the filter obtained after learning is treated selectively with the edge direction after treatment. For example, the middle-most filter in the bottom row is most suitable for edges with a high degree of intensity (90 degree gradient angle) and has a high viscosity (straight rather than curved edges). If the edge of this degree is low, then the other filter is selected as if it were the top line in the graph.
In practice, RAISR will use the filter list once learned to select the most suitable filter for each pixel of the low-resolution picture. When these filters are applied to lower quality images, they will reconstruct details equivalent to the original resolution, which is much better than linear, bicubic, Lancos parsing.
One, background
With the increasing resolution of televisions and moving screens, people’s demand for high-definition pictures has also grown. But this also creates a certain turbulence for the user - to see the big picture in high definition means that it takes up a lot of bandwidth, one is the increase in data costs, and second, the loading speed will slow down, resulting in poor user experience.
The
Therefore, how to achieve high-definition effects through the transmission of small images without affecting the user experience is a very worthwhile discussion. In October 2016, Google published a paper about their introduction of a new technology called Rapid and Accurate Image Super-Resolution (RAISR), which uses machine learning to convert low-resolution images into high-resolution images. This technology can reach resolutions and even exceed the original image at a bandwidth of 75%. At the same time, the speed can be increased by about 10 to 100 times.
Second, demand analysis
function:
1. Low-quality web pictures are converted into high-definition pictures
2. Stop deflation of high-definition pictures on the cloud and reduce the size of pictures
Third, complete the plan
The hyperfraction model construction diagram is as follows:
Google uses machine learning to exercise the program with a pair of low-resolution, high-resolution pictures to find filters that can be selectively applied to each pixel in a low-resolution picture, which can produce details that are comparable to the original picture. There are currently two ways to exercise RAISR:
· The first is a "direct" approach where the filter learns directly in pairs of high and low resolution pictures.
The second approach requires the application of low-power upsampling on low-resolution pictures first, followed by learning filters in the combination of up-sampled pictures and high-resolution pictures.
· The "direct" approach is faster to dispose of, but the second approach takes care of non-integer-wide elements and is better off with hardware performance.
Either way, the RAISR filter is based on the edge features of the image: brightness and color gradients, level and texture areas. This is affected by direction (edge angle), strength (strength with higher sharp edges), and coherence (a measure of the quantification of edge directionality). The following is a set of RAISR filters learned from 10,000 pairs of high and low resolution pictures (low resolution pictures are upsampled). This exercise takes about 1 hour.
Note: 3x super-resolution learning, achieved 11×11 filter convergence. Filters can be learned from a variety of super-resolution elements, including local super-resolution. Note that when the angle of the edge in the figure changes, the filter angle also rotates. Similarly, as the intensity progresses, the sharpness of the filter also improves; as the viscosity progresses, the filter's anisotropy also improves.
From left to right, the filter obtained after learning is treated selectively with the edge direction after treatment. For example, the middle-most filter in the bottom row is most suitable for edges with a high degree of intensity (90 degree gradient angle) and has a high viscosity (straight rather than curved edges). If the edge of this degree is low, then the other filter is selected as if it were the top line in the graph.
In practice, RAISR will use the filter list once learned to select the most suitable filter for each pixel of the low-resolution picture. When these filters are applied to lower quality images, they will reconstruct details equivalent to the original resolution, which is much better than linear, bicubic, Lancos parsing.