CN-121998833-A - Text controllable semantic perception image enhancement method for power equipment inspection

CN121998833ACN 121998833 ACN121998833 ACN 121998833ACN-121998833-A

Abstract

The application relates to a text controllable semantic perception image enhancement method and a text controllable semantic perception image enhancement system for power equipment inspection, which are used for acquiring an image to be enhanced, acquired by a power equipment inspection terminal, and extracting features of the image to be enhanced to obtain image state information. And acquiring text control information corresponding to the inspection task, fusing the text control information with the image state information to generate a joint state representation, inputting the joint state representation into a trained parameter generation model, and outputting a low-dimensional image enhancement parameter vector for controlling the inspection image enhancement process by the parameter generation model. And constructing a parameterized mapping function according to the low-dimensional image enhancement parameter vector, carrying out pixel-level mapping processing on the to-be-enhanced inspection image by using the parameterized mapping function, and outputting an enhanced inspection image for detecting defects or evaluating states of the power equipment as a final enhancement result, thereby being suitable for high-efficiency enhancement processing of the power equipment inspection image in a complex environment.

Inventors

LI KENLI
TANG DING
CHEN XIAOYING
DUAN MINGXING
XIAO GUOQING
XIAO ZHENG
Luo Huizhang
ZHANG JIAPENG
LIU CHUBO

Assignees

湖南大学

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (10)

1. The text controllable semantic perception image enhancement method for the power equipment inspection is characterized by comprising the following steps of: Acquiring an image to be enhanced, which is acquired by an electric power equipment inspection terminal, and extracting features of the image to be enhanced to obtain image state information for representing brightness distribution, contrast statistics and color distribution; Acquiring text control information corresponding to a patrol task, and fusing the text control information with the image state information to generate a joint state representation; inputting the joint state representation into a parameter generation model which is trained, and outputting a low-dimensional image enhancement parameter vector for controlling the inspection image enhancement process by the parameter generation model, wherein the low-dimensional image enhancement parameter vector is used for representing a pixel mapping relation; The parameterized mapping function comprises a lookup table mapping function and/or a curve-based mapping function, and the parameterized mapping function meets brightness mapping monotonicity constraint and edge gradient preserving constraint; and carrying out pixel-level mapping processing on the to-be-enhanced inspection image by using the parameterized mapping function, and outputting an enhanced inspection image as a final enhancement result, wherein the enhanced inspection image is used for detecting defects or evaluating states of the power equipment.
2. The method of claim 1, wherein extracting features from the inspection image to be enhanced to obtain image state information for characterizing brightness distribution, contrast statistics, and color distribution, comprises: performing size normalization and pixel value normalization on the to-be-enhanced inspection image to be preprocessed in a [0,1] interval; Converting the preprocessed image from the RGB color space to a luminance-chrominance separated color space; Image state features are extracted from the converted image, the image state features including at least one of histogram distribution of luminance channels, statistical moment features of chrominance channels, and a global contrast metric.
3. The method of claim 2, wherein the parameter generation model is constructed based on a maximum entropy reinforcement learning framework and comprises a policy network and a value function network, wherein the policy network receives the joint state representation and outputs a conditional probability distribution of the low-dimensional image enhancement parameter vector, and wherein the optimization objective function of the parameter generation model is a weighted sum of a maximum expected cumulative prize and a policy entropy.
4. The method of claim 3, wherein the low-dimensional image enhancement parameter vector comprises at least one of a control point coordinate sequence of Bezier curves for defining a luminance map, a node value vector for populating a one-dimensional luminance look-up table or a three-dimensional color look-up table, and a contrast scaling factor and a saturation bias for adjusting a global attribute.
5. The method of claim 4, wherein the low-dimensional image enhancement parameter vector comprises n+1 control point coordinates when constructing an n-th order bezier curve mapping function Wherein , For the normalized input intensity, the input intensity is calculated, For the normalized intensity value x of any input pixel, the mapped output intensity value y is calculated by the following formula: ; ; Wherein, the Is a Bernstein-based polynomial, Is a binomial coefficient.
6. The method of claim 4, wherein the low-dimensional image enhancement parameter vector defines an output color value for each mesh vertex on a regular cube mesh in RGB or YCbCr color space when constructing a three-dimensional look-up table mapping function, the regular cube mesh dividing L nodes uniformly across each color dimension to form Unit cube element for input color vector The mapping process comprises the following steps: determining the unit cube unit of the input vector, and reading the predefined output color values at 8 vertexes of the unit cube unit Wherein ; Calculating the left lower corner vertex of the input vector relative to the unit cube cell Normalized offset of (a) Calculating final output color values using tri-linear interpolation : ; ; Wherein, the Is an interpolation weight.
7. The method according to claim 1, wherein the pixel-level mapping process is performed on the inspection image to be enhanced by using the parameterized mapping function, and the pixel-level mapping process is implemented by a dedicated hardware acceleration unit, and the hardware acceleration unit includes: a configuration interface for receiving the lookup table data from the software layer; a high-speed on-chip memory for storing the look-up table data; the parallel multipath pixel processing engine is used for carrying out synchronous table look-up and interpolation operation on the input pixel stream; And the output interface is used for outputting the enhanced pixel stream in video time sequence.
8. The method according to any one of claims 1 to 7, further comprising: and in the deployment stage, carrying out feedback evaluation on the enhanced inspection image, and carrying out on-line self-adaptive adjustment on the image enhancement parameters according to a feedback evaluation result under the condition that the parameters of the parameter generation model are not updated, so that the enhancement result meets the preset multi-target constraint condition.
9. The method of claim 8, wherein optimizing the reward signal calculated based on the cross-modal semantic perception model during a training phase of the parameter generation model comprises: Extracting visual feature vectors of the enhanced output image through the image encoder of the cross-modal semantic perception model Extracting text feature vectors of target text semantic instructions through a text encoder ; Computing visual feature vectors And text feature vector Cosine similarity between them as the base value of the bonus signal : ; For the basic value Standardized processing is carried out to obtain a reward signal for training : Wherein And Moving average and standard deviation calculated for historical rewards; updating parameters phi of the strategy network and parameters of the value function network by a soft strategy iterative algorithm using the reward signal r 。
10. A text-controllable semantic-aware image enhancement system for power equipment inspection, comprising: The input module is used for acquiring an image to be enhanced, acquired by the power equipment inspection terminal, and extracting characteristics of the image to be enhanced to obtain image state information for representing brightness distribution, contrast statistics and color distribution; the state characterization and fusion module is used for acquiring text control information corresponding to the inspection task, and fusing the text control information with the image state information to generate a joint state representation; The semantic guided reinforcement learning decision module is used for inputting the joint state representation into a parameter generation model which is trained, and outputting a low-dimensional image enhancement parameter vector used for controlling the inspection image enhancement process by the parameter generation model; The parameterized mapping function construction module is used for constructing a parameterized mapping function according to the low-dimensional image enhancement parameter vector, wherein the parameterized mapping function comprises a lookup table mapping function and/or a curve-based mapping function, and the parameterized mapping function meets brightness mapping monotonicity constraint and edge gradient preserving constraint; the image enhancement execution module is used for carrying out pixel-level mapping processing on the to-be-enhanced inspection image by utilizing the parameterized mapping function and outputting an enhanced inspection image as a final enhancement result, wherein the enhanced inspection image is used for defect detection or state evaluation of the power equipment.

Description

Text controllable semantic perception image enhancement method for power equipment inspection Technical Field The application relates to the technical field of intelligent inspection and computer vision of a power system, in particular to a text controllable semantic perception image enhancement method for inspection of power equipment. Background In the operation and maintenance of a power system, vision-based inspection of power equipment is an important means for finding hidden trouble of the equipment. However, the quality of the inspection image is often affected by complex illumination on site, weather conditions and stains of equipment, so that the defect characteristics are fuzzy, the contrast is insufficient, the inspection is seriously dependent on manual experience, the efficiency is low, and the inspection image cannot be suitable for the inspection of the power equipment in a complex environment. How to realize the efficient enhancement processing of the inspection images of the power equipment in a complex environment is a problem to be solved urgently. Disclosure of Invention Based on the above, it is necessary to provide a text controllable semantic perception image enhancement method, system, device and storage medium for power equipment inspection, which are applicable to power equipment inspection in complex environments. The first aspect of the application provides a text controllable semantic perception image enhancement method for power equipment inspection, comprising the following steps: Acquiring an image to be enhanced, which is acquired by an electric power equipment inspection terminal, and extracting features of the image to be enhanced to obtain image state information for representing brightness distribution, contrast statistics and color distribution; Acquiring text control information corresponding to a patrol task, and fusing the text control information with the image state information to generate a joint state representation; inputting the joint state representation into a parameter generation model which is trained, and outputting a low-dimensional image enhancement parameter vector for controlling the inspection image enhancement process by the parameter generation model, wherein the low-dimensional image enhancement parameter vector is used for representing a pixel mapping relation; The parameterized mapping function comprises a lookup table mapping function and/or a curve-based mapping function, and the parameterized mapping function meets brightness mapping monotonicity constraint and edge gradient preserving constraint; and carrying out pixel-level mapping processing on the to-be-enhanced inspection image by using the parameterized mapping function, and outputting an enhanced inspection image as a final enhancement result, wherein the enhanced inspection image is used for detecting defects or evaluating states of the power equipment. In one embodiment, feature extraction is performed on the to-be-enhanced inspection image to obtain image state information for characterizing brightness distribution, contrast statistics and color distribution, including: performing size normalization and pixel value normalization on the to-be-enhanced inspection image to be preprocessed in a [0,1] interval; Converting the preprocessed image from the RGB color space to a luminance-chrominance separated color space; Image state features are extracted from the converted image, the image state features including at least one of histogram distribution of luminance channels, statistical moment features of chrominance channels, and a global contrast metric. In one embodiment, the parameter generation model is constructed based on a maximum entropy reinforcement learning framework and comprises a strategy network and a value function network, wherein the strategy network receives the joint state representation and outputs conditional probability distribution of the low-dimensional image reinforcement parameter vector, and an optimization objective function of the parameter generation model is a weighted sum of a maximum expected cumulative reward and strategy entropy. In one embodiment, the low-dimensional image enhancement parameter vector comprises at least one of a control point coordinate sequence of a Bezier curve for defining a brightness map, a node value vector for populating a one-dimensional brightness lookup table or a three-dimensional color lookup table, and a contrast scaling factor and a saturation bias for adjusting a global attribute. In one embodiment, when constructing the n-order Bezier curve mapping function, the low-dimensional image enhancement parameter vector contains n+1 control point coordinatesWherein,For the normalized input intensity, the input intensity is calculated,For the normalized intensity value x of any input pixel, the mapped output intensity value y is calculated by the following formula: ; ; Wherein, the Is a Bernstein-based polynomial,Is a binomial c