Search

CN-122023211-A - Low-light image enhancement and self-adaptive dimming method based on semi-supervision

CN122023211ACN 122023211 ACN122023211 ACN 122023211ACN-122023211-A

Abstract

The invention discloses a low-light image enhancement and self-adaptation dimming method based on semi-supervision, which comprises the following steps of constructing a dimming module based on a mixed density network, converting a high-quality image into a low-light image by using the dimming module to generate a combined low-light and normal-light image pair as training data, training an image enhancement model of a UNet framework on the combined training data in an unsupervised mode, fine-tuning by using a small number of real image pairs, controlling a low-light effect by setting a dimming factor in the image enhancement process, and controlling the brightness level of an output image by means of conditional input. The invention has the beneficial effects that the technical scheme realizes multi-dimensional breakthrough in the aspects of data efficiency, sense of reality simulation, enhanced quality and flexibility through a semi-supervised learning strategy, a physically driven dimming module and self-adaptive brightness control, and effectively solves the core pain point of the traditional low light enhancement method.

Inventors

  • DING HAIQIN
  • LI NING
  • Qin Tongchun
  • JIA GUODONG
  • JI GUANG
  • LI ZHIWEI
  • CUI LINGLING

Assignees

  • 南通理工学院
  • 江苏优众微纳半导体科技有限公司

Dates

Publication Date
20260512
Application Date
20251223

Claims (10)

  1. 1. A low-light image enhancement and self-adaptive dimming method based on semi-supervision is characterized by comprising the following steps: S1, constructing a dimming module based on a mixed density network, wherein the dimming module is used for simulating color distortion and dimming processes under low illumination conditions; S2, converting the high-quality image into a low-illumination image by using the dimming module, and generating a synthesized low-light-normal-light image pair as training data; S3, training an image enhancement model of a UNet framework on the synthetic training data in an unsupervised mode, and performing fine adjustment by using a small number of real image pairs; S4, controlling the low illumination effect by setting a dimming factor in the image enhancement process, and controlling the brightness level of the output image by the condition input.
  2. 2. The semi-supervised low light image enhancement and adaptive dimming method according to claim 1, wherein in step S1, the dimming module calculates the ith channel of the reflectivity R by channel-by-channel normalization, and the calculation formula is: Wherein, the And For inputting images And a first reflectance of R The flow path of the liquid is provided with a channel, Representing Hadamard element level segmentation.
  3. 3. The method of claim 1, wherein in step S1, the dimming module adjusts the reflection component R through a mixed density network, the MDN outputs the adjusted reflection component RD according to the inputted reflection component R, the normal illumination component L and the dimmed illumination component LD, and obtains a dark image ID=RD+LD through Hadamard product combination RD and LD.
  4. 4. The semi-supervised based low-light image enhancement and adaptive dimming method as set forth in claim 1, wherein in step S2, the high quality image is derived from a public computer vision dataset, and the resultant image pair generated by the dimming module is used for an unsupervised training phase of the model.
  5. 5. The semi-supervised based low-light image enhancement and adaptive dimming method as set forth in claim 1, wherein the fine tuning process uses the real image pair when constructing the dimming module to adapt the imaging characteristics of the specific camera under low-light conditions in step S3.
  6. 6. The semi-supervised low light image enhancement and adaptive dimming method according to claim 1, wherein the dimming factor is gamma, low light effects with different degrees are generated by adjusting gamma, the condition input is a brightness value delta m, and a UNet architecture is embedded to control the brightness level of an output image.
  7. 7. The semi-supervised low-light image enhancement and adaptive dimming method as set forth in claim 1, wherein the MDN performs independent processing on the 5-dimensional representation of each pixel, and the average value, standard deviation and mixing coefficient of each color channel are output to describe the color distortion probability distribution by sharing weights by the multi-layer perceptron.
  8. 8. The semi-supervised low-light image enhancement and adaptive dimming method as claimed in claim 1, wherein the brightness value Δm is embedded in a manner similar to time embedding in a diffusion model, so that UNet dynamically adapts to target brightness during enhancement.
  9. 9. The method for enhancing and adaptively adjusting luminance of a low-light image based on semi-supervision according to claim 1, wherein in step S3, the input of the UNet model comprises a dimmed image, a dimmed image after histogram equalization, color mapping and illumination information, a residual map U is output, and the final enhanced image is obtained by adding the residual map U and the dimmed image.
  10. 10. The low-light image enhancement and self-adaptive dimming method based on semi-supervision according to claim 1, wherein the semi-supervision is realized by the following steps: generating a large number of synthesized low light-normal light image pairs by using a dimming module, and performing unsupervised pre-training; A small number of real low-light-normal-light image pairs are used to fine tune the UNet model to balance generalization ability with specific scene adaptability.

Description

Low-light image enhancement and self-adaptive dimming method based on semi-supervision Technical Field The invention relates to the field of computer vision and image processing, in particular to a low-light image enhancement and self-adaptive dimming method based on semi-supervision. Background The low-light image enhancement is a key technology in the fields of computer vision and image processing, and is widely applied to scenes such as night monitoring, mobile phone photography, automatic driving vision perception and the like. The method has the core challenges of keeping the detail information of the original scene, suppressing noise and maintaining natural color distribution while improving the brightness of the image. However, the prior art still has the following significant limitations in practical applications: The traditional full supervision method needs to rely on a large number of paired low-light-normal-light image data (such as LOL data sets) for training, and the model is excellent in specific data sets, but when facing different camera equipment, scenes or illumination conditions, the generalization capability is obviously reduced due to imaging characteristic differences (such as sensor noise and lens distortion). The unsupervised method gets rid of the dependence on the paired data, and enhances the paired data by image decomposition (such as Retinex theory) or generation of a countermeasure network (GAN), but the unsupervised method often has problems of overexposure, color distortion or detail blurring, and the like, so that the brightness improvement and the naturalness are difficult to balance. The existing method is mostly based on synthetic low-light data training, and has distribution difference with a real low-light environment. For example, the composite data tends to ignore nonlinear responses of camera hardware under low light (e.g., noise due to ISO gain, white balance shift), resulting in a model that is prone to color bias or artifacts when processing truly taken low light images. In addition, the low-light imaging characteristic difference of cameras of different brands and models further aggravates the adaptation difficulty of the model. Most enhancement algorithms adopt a fixed target brightness strategy, and cannot meet the personalized brightness requirement of a user on an output image (such as adjusting brightness according to a scene). Some methods attempt to introduce brightness adjustment parameters, but are mostly implemented by post-processing, which easily breaks the overall contrast and color consistency of the image. Although semi-supervised learning has potential in reducing data dependence, in the existing low-light enhancement method, the design of a semi-supervised strategy is still immature, or only a small amount of real data and synthetic data are simply mixed, physical characteristics of low-light imaging are not modeled, or fine processing on key factors such as reflectivity, illumination components and the like is lacking, so that the enhancement effect is different from that of a full-supervision method. Aiming at the problems, the invention provides a semi-supervision method combining a mixed density network dimming module and a conditional UNet architecture, which aims to realize double breakthrough of data efficiency and enhanced quality by simulating a real low-light imaging process and dynamically controlling brightness output. Disclosure of Invention The invention mainly solves the technical problem of providing a semi-supervision-based low-light image enhancement and self-adaptive dimming method, which solves one or more of the problems in the prior art. In order to solve the technical problems, the invention adopts a technical scheme that the low-light image enhancement and self-adaptive dimming method based on semi-supervision is characterized by comprising the following steps: S1, constructing a dimming module based on a mixed density network, wherein the dimming module is used for simulating color distortion and dimming processes under low illumination conditions; S2, converting the high-quality image into a low-illumination image by using the dimming module, and generating a synthesized low-light-normal-light image pair as training data; S3, training an image enhancement model of a UNet framework on the synthetic training data in an unsupervised mode, and performing fine adjustment by using a small number of real image pairs; S4, controlling the low illumination effect by setting a dimming factor in the image enhancement process, and controlling the brightness level of the output image by the condition input. In some embodiments, in step S1, the dimming module calculates the ith channel of the reflectivity R by channel-by-channel normalization, where the calculation formula is: Wherein, the AndFor inputting imagesAnd a first reflectance of RThe flow path of the liquid is provided with a channel,Representing Hadamard element level segm