Self-adaptive video enhancement method and device

Abstract

The invention discloses a self-adaptive video enhancement method and device. The method comprises the following steps: acquiring target video data; the target video data is input into a pre-trained video enhancement model, an enhanced video is obtained, and the training sample set is obtained, specifically, a generative adversarial network is constructed, the generative adversarial network comprises a generator and a discriminator, the generator takes the real high-quality video and the degradation features as input to generate a simulated low-quality video, and the discriminator is used for discriminating the real high-quality video; the discriminator is used for judging the fitting degree between the simulated low-quality video and the real low-quality video; training the generative adversarial network to enable the fitting degree between the simulated low-quality video and the real low-quality video to meet a set loss standard; and taking the trained generator as a video degradation learning model so as to generate low-quality videos with different degradation distribution characteristics for the collected high-quality videos, and further constructing the training sample set. According to the invention, the visual definition and fidelity of the target video can be improved.

Description

technical field [0001] The present invention relates to the technical field of image processing, and more particularly, to an adaptive video enhancement method and device. Background technique [0002] The video enhancement method is widely used. Taking laparoscopic surgery as an example, it is a newly developed minimally invasive method. It has the advantages of small trauma, enlarged surgical field, low damage to surrounding tissues, light postoperative wound pain, beautiful appearance, and quick recovery. With the advantages of fewer complications, fewer days of hospitalization, and low cost burden, video enhancement for endoscopic surgery scenes is conducive to improving the efficiency of diagnosis and treatment. However, in the current endoscopic surgery, the image sensor often produces smoke or fog in the patient due to temperature difference or surgical cauterization, which leads to the degradation of the captured image or video quality. Traditional strategies such as...

AI Technical Summary

Problems solved by technology

1) The factors affecting the video quality of endoscopic surgery are complex, and the image degradation process is difficult to model

[0005] 2) Different from video enhancement tasks in general natural scenes, image and video enhancement in the medical field requires higher fidelity, and doctors cannot accept poor contrast, color distortion, blur, light, etc. in the surgical field of view due to video enhancement. Problems such as halo artifacts, let alone receiving non-existent content during the enhancement process, these problems will greatly affect the doctor's visual judgment and have an adverse effect on the surgical process, such as the cyan lightening or changing of blood vessels, which will cause doctors to misjudge Therefore, it is possible to accidentally injure the blood vessel; shadows appear in places where there are no shadows, which may make doctors misjudge the scope of operation

Therefore, the image enhancement process requires high fidelity, and it is necessary to avoid introducing false details and content that do not exist due to algorithm enhancement, which will affect the doctor's judgment

[0006] 3) Surgery has high requirements for real-time operation, and excessive delay will affect the operation and bring medical risks

[0007] To sum up, for endoscopic video data, it is difficult to obtain high-quality-low-quality data pairs, but the requirements for enhanced details and real-time performance are higher, which makes the existing methods unable to achieve ideal results. Therefore, it is necessary to optimize the model structure. , loss function and training strategy are designed more finely to improve the effect of video enhancement

Images