Image Enhancement Model With Brightness-Interval Convolution Branches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image enhancement technologies fail to effectively improve both brightness and contrast simultaneously, particularly in low-light conditions, leading to inadequate brightening or excessive brightness changes in night-scene images.
Innovation Solution
A neural network-based image enhancement model with a Histogram Consistency Module (HCM) that includes convolution branches corresponding to preset brightness intervals, applying specific convolution processing to enhance brightness and contrast while maintaining low-brightness areas unchanged, utilizing alignment and fusion modules to combine multiple frames for reduced noise and enhanced details.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image enhancement methods (histogram equalization, contrast limited adaptive histogram equalization) are used, then processing speed is fast, but image quality and naturalness are poor
Solution Approach 1:
The patent pre-trains multiple image enhancement models with different image qualities before actual use. When enhancing an image, the system selects and applies the most appropriate pre-trained model based on the input image characteristics, avoiding the need to train models in real-time and thus reducing processing time while maintaining high image quality.
Solution Approach 2:
The patent changes the parameter of using multiple models with different image qualities instead of relying on a single model. By selecting different pre-trained models based on input image characteristics, the system optimizes the balance between processing speed and image quality dynamically.
2Measurement precision
If deep learning-based image enhancement models are trained from scratch, then image quality can be improved, but training time and computational resources are excessive
Solution Approach 1:
The patent performs image enhancement model training in advance (pre-training) to generate multiple models with different image qualities. This preliminary action transfers the time-consuming training process to the model preparation phase, allowing fast inference during actual image enhancement while maintaining high image quality.
Solution Approach 2:
The patent creates multiple copies of image enhancement models with different training levels and image qualities. Instead of training one model from scratch during use, the system selects from pre-created model copies, significantly reducing the time required for image enhancement while maintaining high quality.
3Adaptability or versatility
If a single image enhancement model is used, then the system is simple, but it cannot adapt to different image types and qualities
Solution Approach 1:
The patent creates a universal image enhancement system that can handle multiple image types and qualities by using multiple pre-trained models. Each model is trained on different datasets or with different parameters, enabling the system to adapt to various image characteristics (portraits, landscapes, low-light, etc.) while maintaining a relatively simple selection and application process.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The present disclosure provides a method for training an image enhancement model. The image enhancement model comprises an enhancement module configured to enhance brightness and contrast. The enhancement module comprises convolution branches corresponding to brightness intervals, and is used for inputting pixels of images inputted therein into the corresponding convolution branches according to the brightness intervals to which the pixels belong; convolution is performed by a first convolution unit in each convolution branch; and the images outputted by the convolution branches are merged and then convolved by a second convolution unit. The method comprises: inputting a sample image into the image enhancement model to obtain a result image outputted by the image enhancement model; calculating a loss, the loss comprising an image loss of the result image with respect to a standard image, and a first constraint loss of brightness histogram constraint of the images outputted by the convolution branches with respect to the standard image in each convolution branch; adjusting the enhancement module according to the loss; and if a training end condition is not satisfied, returning to the step of inputting the sample image into the image enhancement model. The present disclosure further provides an image enhancement method and a computer readable medium.