CN-120410864-B - Diffusion road image enhancement method based on wavelet transformation and priori

CN120410864BCN 120410864 BCN120410864 BCN 120410864BCN-120410864-B

Abstract

The invention discloses a diffusion road image enhancement method based on wavelet transformation and priori, which comprises the steps of constructing a training set, constructing a model according to parameters defined by configuration files, putting the model on the training set for training, calculating loss between an output value and a true value of the model according to a loss function, carrying out back propagation on the calculated loss, updating and optimizing parameters of the model, stopping training until the prediction performance reaches a preset value, constructing the model according to model parameters stored in a training stage, putting the model on a test set for testing, and carrying out quantization index calculation on a test result and the true image. The invention combines the highway image enhancement method with the wavelet transformation, the prior graph and the diffusion model, reduces the calculation cost through the wavelet transformation, and simultaneously gives consideration to the generation of high-frequency information by the rich low-frequency information in the prior graph and the diffusion model, so that the model can effectively solve the problem of losing details of low-frequency offset color and high-frequency texture, and obtain excellent visual results.

Inventors

ZHANG YUN
ZHOU ZHOU
GUO HUA
CHENG XIN
SUN SHOUWEI
ZHOU JINGMEI
ZHANG MENG
MIN HAIGEN
ZHOU LINYING

Assignees

长安大学
云南省交通科学研究院有限公司
云南云岭高速公路交通科技有限公司

Dates

Publication Date: 20260512
Application Date: 20250324

Claims (4)

1. The diffusion road image enhancement method based on wavelet transformation and priori is characterized by comprising the following steps: Step A, selecting training data and constructing a training set; Step B, constructing a model according to the parameters defined by the configuration file; step C, training the model by using a training set, calculating the loss between the output value and the true value of the model according to a defined loss function, carrying out back propagation on the calculated loss, and updating and optimizing parameters of the model; The step C comprises the following steps: step C1, random from 1..a time T is taken in T for training, t=1000, random two noise levels obeying the standard normal distribution N (0, 1) And The method is used for noise adding processing of a diffusion model forward process; step C2, the highway image is displayed Wavelet transform to divide into low-frequency sub-bands of road image And high frequency sub-bands Image of the ground Wavelet transform to divide into low-frequency sub-bands of ground image And high frequency sub-bands ; Step C3, low frequency sub-band of ground image Performing noise adding operation to obtain a noise added image ; Step C4, the low frequency sub-band of the highway image Transmittance a priori map m1 corresponding to road image and noise-added image Spliced together and input into a neural network The inner part of the inner part is provided with a plurality of grooves, A finger Unet neural network; The output of the neural network is the predicted noise, and the added noise The parameters of the network are updated by taking a loss function; Step C5, high frequency subband of ground image Performing noise adding operation to obtain a noise added image ; Step C6, high-frequency sub-band of road image Transmittance a priori map m1 corresponding to road image and noise-added image Spliced together and input into a neural network The inner part of the inner part is provided with a plurality of grooves, A finger Unet neural network; the output of the neural network is the predicted noise, and the added noise The parameters of the network are updated by taking a loss function; step C7, randomly generating a noise figure , These two noise patterns obey a standard normal distribution N (0, 1), defining a time step i from T to 1, t=1000, i being taken from 1000 to 1, when i is cycled to 1, will be , Two variables are defined as 0, otherwise , Defined as two random noises, both obeying a standard normal distribution N (0, 1); step C8 for each time step i, Obtained by the following formula: obtained by the following formula: Wherein, the Is constant, each time step i has a corresponding , Is that The image of the last moment in time is displayed, Is that Is from To the point of Is used for the order of the steps of (a), , Is a neural network in the forward direction process, Is the variance of the values, Is a constant; Step C9, circularly executing denoising operation by using the previous image Obtaining Through again Obtaining Until it is obtained Obtaining Similarly, return the noise reduction result 、 Will be And low frequency subbands of ground images Doing a loss function, updating network parameters, and And high frequency sub-bands of ground images A loss function is made, and network parameters are updated; step C10, saving a parameter file of the model trained in the training process for testing; Step D, constructing a model according to model parameters stored in the training stage; And E, placing the model on a test set for testing, and calculating a quantization index between a test result and a real image to measure the performance of the model through the quantization index.
2. The method for enhancing a diffuse road image based on wavelet transform and prior as claimed in claim 1, wherein in said step A, the training data comprises a road image and its corresponding ground real image, and a corresponding transmittance prior image and texture prior image, wherein the transmittance prior image is determined according to the red channel degradation region of the road image, and the texture prior image is calculated by an edge detection operator; the training set comprises six parts ① highway images ② Ground image The image processing system comprises a transmission priori map m1 corresponding to ③ highway images, a texture priori map m2 corresponding to ④ highway images, a transmission priori map m3 corresponding to ⑤ ground images and a texture priori map m4 corresponding to ⑥ ground images.
3. The method for enhancing a diffuse road image based on wavelet transform and prior according to claim 1, wherein in said step B, the parameters of the configuration file comprise input dimension, output dimension, number of channels of the middle layer, and number of residual blocks.
4. The method for enhancing a diffuse road image based on wavelet transform and prior according to claim 1, wherein said step E comprises the steps of: step E1, randomly generating a noise figure , These two noise patterns obey a standard normal distribution N (0, 1), defining a time step i from T to 1, t=1000, i being taken from 1000 to 1, when i is cycled to 1, will be , Two variables are defined as 0, otherwise , Defined as two random noises, both obeying a standard normal distribution N (0, 1); step E2, for each time step i, Obtained by the following formula: obtained by the following formula: Wherein, the Is constant, each time step i has a corresponding , Is that The image of the last moment in time is displayed, Is that Is from To the point of Is used for the order of the steps of (a), , Is a neural network in the forward direction process, Is the variance of the values, Is a constant; step E3, circularly executing denoising operation by using the previous image Obtaining Through again Obtaining Until it is obtained Obtaining Similarly, return the noise reduction result 、 ; Step E4, and then , And inputting the images into an inverse wavelet module, and performing inverse wavelet processing to obtain enhanced images.

Description

Diffusion road image enhancement method based on wavelet transformation and priori Technical Field The invention relates to the field of computer vision and image enhancement, in particular to a diffusion road image enhancement method based on wavelet transformation and priori. Background In recent years, with the development of intelligent traffic systems, acquisition and processing of road images are increasingly emphasized. Road images play an important role in traffic monitoring, accident detection, and vehicle identification applications. However, these images often suffer from blurring, low contrast, and color distortion due to environmental factors, illumination variations, and limitations of the imaging device. These challenges make it difficult for conventional image processing methods to meet practical requirements. The wavelet transformation is used as an effective image processing technology, can analyze the local characteristics of images on different scales, has better time-frequency localization capability, and is suitable for processing images with multi-scale and multi-frequency characteristics. Therefore, it is widely used in image enhancement tasks to improve the sharpness and contrast of images. Meanwhile, the introduction of priori knowledge can further improve the image enhancement effect. By establishing a priori model with physical significance, the characteristics of the road image, especially the image degradation in a complex environment, can be better adapted. The diffusion method combining wavelet transformation and priori knowledge can effectively cope with the diversity and uncertainty existing in the road image, thereby improving the image quality and meeting the high standard requirements of traffic management and monitoring. In summary, the diffusion road image enhancement method based on wavelet transformation and priori knowledge provides effective technical support for improving the definition, contrast and color reproducibility of road images, and promotes the further development of intelligent traffic systems. In recent years, road surface image enhancement methods can be classified into the following three types: 1. road surface image enhancement based on an image processing method; 2. restoring road surface images based on the degradation model; 3. road surface image enhancement based on deep learning. The image processing-based road surface image enhancement method aims at improving the visual effect of a road surface image by adjusting single-point pixel values. Typically, these methods employ two input channels generated from degraded images and adjust the image at the detail, structure and illumination level based on the different sparse representation capabilities of the square norms of the image space information to improve the details and contrast of the image. Although these methods have made some progress in improving the image quality of the road surface, there are some limitations. First, image processing-based methods focus mainly on the adjustment of individual pixels, and it is difficult to efficiently capture and connect global information of images. In the road environment, global information is critical to restore the appearance of real scenes and objects. Due to such limitations of local operation, the method may lead to problems of excessive or insufficient image enhancement, especially in the presence of complex illumination and scattering. Second, since the road surface image is affected by strong scattering and absorption of light, details and structures in the image are lost or blurred. Image processing-based methods often have difficulty effectively handling such complex optical distortions, and therefore present challenges in restoring the details and structure of the image. The degradation model-based road surface image enhancement method aims at realizing the inverse solution of an undegraded road surface image by modeling the specific degradation process of the road image. First, complex physical processes such as light propagation and scattering in the road environment make it difficult to accurately model the degradation process of the image. The quality of road images is affected in many ways by phenomena such as illumination variation, weather factors and vehicle types, and these effects show differences under different road conditions. Therefore, when image degradation modeling is performed based on a specific model, it is often difficult to fully consider the complexity of the road environment, resulting in a limitation in the generalization capability of the method. Secondly, due to the diversity of road image acquisition conditions, such as different illumination, weather, traffic density and other factors, the robustness of the method based on the specific model in different scenes is poor. This makes it possible for these methods to exhibit large performance fluctuations in practical applications, and it is difficult to obtain st