Image enhancement method for detecting security and protection equipment of working personnel

Abstract

The invention belongs to the field of image enhancement, and particularly relates to an image enhancement method for security equipment detection of workers. The method comprises the following steps: firstly, collecting an image containing security and protection equipment and preprocessing; fusing the regional brightness features and multi-modal data to construct an input vector, and realizing improved dynamic brightness compensation by means of a neural network and reinforcement learning; then, improving adaptive frequency domain multi-scale CLAHE, and covering an adaptive local window, dynamic histogram cutting and multi-scale fusion CLAHE; and finally, detecting the security equipment by using the enhanced image. According to the method, excessive enhancement or information loss of a traditional method can be avoided, details of different areas are accurately enhanced, the overall and local contrast ratio of the image is optimized, the identifiability of security equipment in a complex environment is improved, adaptive optimization can be achieved during detection, the enhancement effect is stable and reliable, and the security equipment detection problem in an actual scene is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and particularly relates to an image enhancement method for detecting security equipment of staff members. Background Art

[0002] In today's security work, staff members highly rely on security equipment. It is crucial to accurately and timely detect whether staff members correctly wear and use security equipment. However, image acquisition in actual scenarios faces many challenges, resulting in uneven image quality, which brings great difficulties to the detection of security equipment. In different working environments, the light intensity and angle vary significantly. When outdoors with direct strong light, the image is prone to over-bright areas, causing some details of the security equipment to be lost; while in indoor dim light or night scenes, the image will be too dark to clearly distinguish the features of the equipment. This uncertainty of light greatly reduces the accuracy of image-based security equipment detection. The working scene is often full of various interference factors. There may be objects in the background that are similar in color and shape to the security equipment, or there may be occlusions that partially or completely block the security equipment. In addition, the diversity of the dressing styles, postures, and movements of different staff members also increases the complexity of the image, further increasing the difficulty of accurately detecting the security equipment. Existing image enhancement methods have limitations. For example, the simple histogram equalization method can, to a certain extent, improve the overall brightness and contrast of the image, but it is prone to over-enhancing or losing image details, resulting in image distortion. Some brightness compensation methods based on global adjustment cannot make targeted adjustments according to the characteristics of different regions in the image and are difficult to meet the requirements of accurately detecting security equipment under complex backgrounds and lighting conditions. Traditional frequency-domain processing methods lack adaptability in window selection and contrast adjustment and cannot effectively handle the differences in different images. Summary of the Invention

[0003] The present invention aims at the technical problems existing in the above background art and proposes an image enhancement method for detecting security equipment of staff members.

[0004] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. First, collect image information containing the security equipment of staff members and preprocess the image;

[0006] S2. Then, fuse the regional brightness features and multi-modal data to construct an input vector and implement improved dynamic adjustment of image brightness compensation;

[0007] S3. Improve the adaptive frequency-domain multi-scale CLAHE, including improving the adaptive local window of the frequency-domain energy distribution, improving the dynamic histogram clipping of local contrast perception, and improving the multi-scale fusion CLAHE;

[0008] Among them, for the improvement of the adaptive local window of the frequency-domain energy distribution, by analyzing the energy distribution of different regions of the image in the frequency domain, the size of different local windows is adaptively selected. First, calculate the frequency-domain energy distribution within the local region: where F(I comp (x,y)) is the Fourier transform of the image at the position (x,y), H high (u,v) is the high-pass filter, and E f (x,y) represents the local high-frequency energy. Then, perform adaptive window selection: where S max is the maximum window size, S min is the minimum window size, and E max is the maximum high-frequency energy of the entire image;

[0009] Among them, for the improvement of the dynamic histogram clipping of local contrast perception, by calculating the local contrast, the histogram clipping threshold is dynamically adjusted for adaptive clipping. The adjustment of the adaptive clipping threshold is: where C max , C min are the maximum and minimum clipping limits respectively, and C local (x,y) is the local contrast;

[0010] Among them, for the improvement of the multi-scale fusion CLAHE, CLAHE enhancement is performed through three different window scales of small, medium, and large, and weighted fusion is carried out. First, equalize the image: where S s represents the window size of different scales, and s ∈ {1, 2, 3} represents three different scales of small, medium, and large. Then, fuse the final enhanced image according to the adaptive fusion coefficient: where ω s (x,y) is the adaptive fusion coefficient;

[0011] S4. Finally, use the enhanced image to detect the security equipment of the staff to ensure accurate recognition under different lighting conditions and complex environments.

[0012] Preferably, in step S2, the construction of the input vector is to obtain the average brightness based on regional brightness analysis and the standard deviation of brightness and add the image color feature vector m represents the dimension of the color feature vector, the depth feature D i , and the overall contrast of the image is C g, then the input vector X of the neural network is: where N represents different divided regions.

[0013] Preferably, the average brightness where |R i | represents the number of pixels in region R i , and the standard deviation of brightness

[0014] Preferably, the implementation of dynamically adjusting the image brightness compensation in step S2 is to output through the neural network based on the obtained input vector X, and the output is the brightness compensation parameter vector K = [k 1 , k 2 , k 3 ,..., k N for each region. The j-th output k j of the neural network is expressed as: where α ji is the weight connecting the i-th neuron in the input layer and the j-th neuron in the output layer, and b j is the bias of the j-th neuron in the output layer; among them, when introducing reinforcement learning, the weights and biases are dynamically adjusted according to the reward feedback, and the weights and biases are updated using the method of stochastic gradient. When the security equipment in the image undergoes dynamic changes and triggers the re - division of regions, the features of each region are recalculated and the compensation parameters are adjusted. The compensation parameters of the new region are obtained by the updated neural network according to the newly updated input vector, and the compensated image is where are the corresponding new compensation parameters, is the new region.

[0015] Preferably, the specific form of the adaptive fusion coefficient ω s (x, y) is: where C s represents the global contrast of different scales.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: an input vector is constructed by using regional brightness features and multi-modal data, the brightness compensation parameters are calculated in combination with a neural network, and continuously optimized through reinforcement learning, so that the image can maintain the best brightness under different lighting conditions, avoiding over-enhancement or information loss caused by traditional global adjustment methods. By analyzing the frequency-domain energy distribution of local regions of the image, the local window size is adaptively selected, so that different regions are subjected to targeted enhancement processing, improving the detail expressiveness and effectively reducing the influence of noise. According to the local contrast characteristics, the histogram clipping threshold is adaptively adjusted to ensure that the contrast enhancement of the image in different regions is more accurate, while avoiding the loss of detail information due to over-equalization and improving the image clarity. The CLAHE enhancement results of three different window scales of small, medium, and large are combined, and weighted by using an adaptive fusion coefficient, so that the overall details and local contrast of the image are optimized, improving the identifiability of security equipment in complex backgrounds. Reinforcement learning is introduced during the detection process. When the security equipment in the image is adjusted due to changes in posture, lighting, or environment, the system can dynamically update the neural network weights to achieve adaptive optimization, making the enhancement effect still stable and reliable during long-term operation. Detailed implementation manners

[0017] In order to more clearly understand the above objects, features, and advantages of the present invention, the present invention will be further described below with reference to the embodiments. It should be noted that the embodiments of the present application and the features in the embodiments may be combined with each other without conflict.

[0018] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0019] Embodiment. In today's security work, the staff highly rely on security equipment. It is crucial to accurately and timely detect whether the staff correctly wear and use the security equipment. However, image acquisition in the actual scenario faces many challenges, resulting in uneven image quality, which brings great difficulties to the detection of security equipment. Therefore, an image enhancement method for the detection of security equipment for staff is proposed.

[0020] First, in order to ensure the image quality, image information containing security equipment for staff is first collected, and the image is preprocessed. The preprocessing steps include operations such as denoising, normalization, and image alignment to ensure that the images for subsequent processing have consistent quality and format.

[0021] Considering the problem of uneven image brightness under different lighting conditions, existing dynamic brightness compensation methods mainly rely on global adjustment strategies, such as histogram equalization, gamma correction, or simple linear transformation. These methods have significant defects when dealing with uneven brightness images, mainly manifested as the inability to perform adaptive adjustment for local regions, resulting in over-enhancement or information loss in some regions of the image. And they fail to combine the content information of the image and only rely on fixed rules for adjustment, making it difficult to adapt to complex environments, such as scenes with strong light, shadows, or interference from multiple light sources. However, the improvement of this strategy is to fuse regional brightness features and multi-modal data, and use a neural network to adaptively optimize the brightness compensation parameters. Therefore, fuse regional brightness features and multi-modal data to construct the input vector, and realize the improved dynamic adjustment of image brightness compensation. The construction of the input vector is to obtain the average brightness based on regional brightness analysis and the standard deviation of brightness Add the image color feature vector m represents the dimension of the color feature vector, the depth feature D i , the overall contrast of the image is C g , then the input vector X of the neural network is:

[0022] where N represents different divided regions. Among them, the average brightness where |R i | represents the number of pixels in region R i , the standard deviation of brightness

[0023] In addition, the implementation of dynamically adjusting the image brightness compensation is to output according to the obtained input vector X through the neural network. The output is the brightness compensation parameter vector K = [k 1 , k 2 , k 3 ,..., k N for each region. The j-th output k j of the neural network is expressed as: where α ji is the weight connecting the i-th neuron in the input layer and the j-th neuron in the output layer, and b j is the bias of the j-th neuron in the output layer; among them, when introducing reinforcement learning, the weights and biases are dynamically adjusted according to the reward feedback, and the weights and biases are updated using the method of stochastic gradient. When the security equipment in the image undergoes dynamic changes and triggers the re-division of regions, the features of each region are recalculated and the compensation parameters are adjusted. The compensation parameters of the new region are obtained by the updated neural network according to the newly updated input vector. The compensated image is where is the corresponding new compensation parameter, For a new area. The advantage of this improvement is that it can adaptively adjust to the brightness differences in different areas, thereby making the illumination balanced, while avoiding the problems of over-enhancement and information loss. In addition, it can continuously optimize the compensation parameters in a dynamically changing environment, enabling the system to maintain an efficient brightness adjustment ability when facing changes in illumination conditions.

[0024] Then, considering that the existing CLAHE has a fixed local window, a statically set contrast clipping threshold, and a single-scale processing method, resulting in poor performance in complex illumination environments and being unable to adaptively adjust to the different regional characteristics of images, it may over-enhance noise in areas with rich high-frequency details, while the enhancement effect is insufficient in low-frequency areas. In the present invention, an adaptive local window for improving the frequency-domain energy distribution, a dynamic histogram clipping for improving local contrast perception, and an improved multi-scale fusion CLAHE are proposed. Among them, for the adaptive local window for improving the frequency-domain energy distribution, by analyzing the energy distribution of different regions of the image in the frequency domain, different local window sizes are adaptively selected. First, calculate the frequency-domain energy distribution within the local region: Where F(I comp (x,y)) is the Fourier transform of the image at the position (x,y), H high (u,v) is a high-pass filter, E f (x,y) represents the local high-frequency energy, and then perform adaptive window selection: Where S max is the maximum window size, S min is the minimum window size, and E max is the maximum high-frequency energy of the entire image. Among them, the dynamic histogram clipping for improving local contrast perception is to calculate the local contrast and dynamically adjust the histogram clipping threshold for adaptive clipping. The adjustment of the adaptive clipping threshold is: Where C max , C min are the maximum and minimum clipping limits respectively, and C local (x,y) is the local contrast. Among them, the improved multi-scale fusion CLAHE is to perform CLAHE enhancement through three different window scales of small, medium, and large, and perform weighted fusion. First, equalize the image: Where S s represents the window size of different scales, s∈{1,2,3} represents three different scales of small, medium, and large. Then, fuse the final enhanced image according to the adaptive fusion coefficient: Where ω s (x,y) is the adaptive fusion coefficient, Where C s represents the global contrast of different scales.

[0025] Finally, the enhanced image is used to detect the security equipment of the staff to ensure accurate recognition under different lighting conditions and complex environments.

[0026] The above are only the preferred embodiments of the present invention, and are not limitations to the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An image enhancement method for detecting staff security equipment, characterized in that: The following steps are involved: S1. First, collect image information containing staff security equipment and pre-process the image; S2, then fuse the regional brightness features and multimodal data to construct the input vector, and realize improved dynamic adjustment of image brightness compensation; S3, improved adaptive frequency domain multi-scale CLAHE, including improved adaptive local window for frequency domain energy distribution, improved dynamic histogram clipping for local contrast perception, and improved multi-scale fusion CLAHE; The adaptive local window of frequency domain energy distribution is improved. The energy distribution of different regions of the image is analyzed in the frequency domain, and different local window sizes are adaptively selected. The frequency domain energy distribution in the local region is first calculated: Among them, F(I comp (x,y)) is the Fourier transform of the image at position (x,y), H high (u,v) is a high-pass filter, E f (x,y) represents the local high-frequency energy, and then adaptive window selection is performed: Where S max is the maximum window size, S min is the minimum window size, E max is the maximum high-frequency energy of the entire image; The dynamic histogram clipping that improves local contrast perception is to calculate the local contrast and dynamically adjust the histogram clipping threshold for adaptive clipping. The adaptive clipping threshold is adjusted as follows: Among them C max ,C min are the maximum and minimum clipping limits, C local (x,y) is the local contrast; The improved multi-scale fusion CLAHE uses three different window scales of small, medium and large to perform CLAHE enhancement and weighted fusion. First, the image is equalized: Where S s Represents the window size of different scales, s∈{1,2,3} represents three different scales of small, medium and large, and then the final enhanced image is fused according to the adaptive fusion coefficient: where ω s (x,y) is the adaptive fusion coefficient; S4. Finally, the enhanced image is used to detect the staff's security equipment to ensure accurate identification under different lighting conditions and complex environments.

2. The image enhancement method for detecting staff security equipment according to claim 1, characterized in that: The input vector in step S2 is constructed by obtaining the average brightness based on regional brightness analysis. and brightness standard deviation Increase the image color feature vector m represents the dimension of the color feature vector, and the depth feature D i , the overall image contrast is C g , then the input vector X of the neural network is: Where N represents different divided areas.

3. The image enhancement method for detecting staff security equipment according to claim 2, characterized in that: Average brightness Where |R i | indicates region R i The number of pixels, brightness standard deviation 4. The image enhancement method for detecting staff security equipment according to claim 1, characterized in that: The implementation of dynamically adjusting the image brightness compensation in step S2 is to output the obtained input vector X through a neural network, and the output is a brightness compensation parameter vector K = [k1, k2, k3, ..., k N ], the jth output k of the neural network j It is expressed as: where α ji is the weight connecting the i-th neuron in the input layer and the j-th neuron in the output layer, b j is the bias of the jth neuron in the output layer; where reinforcement learning is introduced, weights and biases are dynamically adjusted according to reward feedback, and the weights and biases are updated using the stochastic gradient method. When the security equipment in the image changes dynamically and triggers the redivision of the region, the features of each region are recalculated and the compensation parameters are adjusted. The compensation parameters of the new region are updated by the updated neural network according to the new updated input vector. The compensated image is in is the corresponding new compensation parameter, For new areas.

5. The image enhancement method for detecting staff security equipment according to claim 1, characterized in that: The adaptive fusion coefficient ω in step S3 s The specific form of (x,y) is: Among them C s Represents the global contrast at different scales.

Images