Perceptually driven augmentation network method and system for markerless microscopic cell images

Abstract

The application discloses a perception-driven enhancement network method and system for a label-free microscopic cell image, and relates to the technical field of optical image processing. The method comprises the following steps: through hybrid adaptive Retinex, noise region adaptive detection and partition denoising are performed on an input image, and combined with multi-scale Retinex enhancement, edge compensation and S-shaped curve stretching, an auxiliary input image is generated; the illumination component and the reflection component of the image are estimated; through multi-scale illumination decomposition, the illumination sensitivity coefficient and the local noise level are extracted, and the Gamma correction parameter matched with the local illumination and noise characteristics is generated from the dynamic Gamma parameter; according to the noise level, the illumination component is subjected to partition Gamma correction, and a noise secondary correction image component is obtained; the correction component and the reflection component are multiplied element by element, and are weighted and fused with the auxiliary input image, and finally, the label-free microscopic cell image after enhancement is output. Through the multi-stage cooperation of the perception-driven, the noise is effectively inhibited while the cell structure details are retained, and the image quality is improved.

Description

Technical Field

[0001] This invention relates to the field of optical image processing technology, and more specifically to a perception-driven enhancement network method and system for label-free microscopic cell images. Background Technology

[0002] Optical microscopy, as a fundamental observational technique, has wide applications in cell biology, life sciences, and clinical medical diagnostics. Label-free microscopy, with its non-invasive nature and ability to maintain the original physiological state of cells, has become an important tool in biomedical research. This type of technology performs imaging analysis based on the optical or physical properties of the sample itself—such as refractive index, light scattering, phase distribution, or amplitude information. It not only avoids the complex sample preprocessing procedures of traditional labeling methods but also eliminates side effects such as phototoxicity and fluorescence bleaching, thus demonstrating irreplaceable advantages in dynamic live cell observation and long-term real-time monitoring. Currently, common label-free imaging techniques mainly include bright-field imaging (BF), differential interferometry (DIC), phase contrast imaging (PHC), and Huffman modulation contrast imaging (HMC).

[0003] Despite the advantages mentioned above, label-free microscopic images generally suffer from drawbacks such as overall low brightness, insufficient distinction between target and background, and blurred cell outlines due to the lack of external staining or labeling to enhance imaging contrast. In low-light imaging environments, photon noise and image sensor noise superimpose each other, further weakening the image signal-to-noise ratio and making it difficult to clearly distinguish subcellular structures, severely limiting the accurate identification of cell morphology and subsequent quantitative analysis. Furthermore, variables such as the optical characteristics of the imaging system itself, illumination conditions, equipment parameter configuration, and operator experience can also lead to problems such as loss of local details, weakened edges, and decreased contrast, seriously affecting the performance of automated image analysis tasks such as cell segmentation, detection, and counting.

[0004] A key bottleneck in this technological field has long been the severe lack of high-quality paired training data. Because label-free imaging does not rely on exogenous contrast agents, it is impossible to simultaneously acquire "low-quality-high-quality" image pairs with significant quality differences within the same field of view. This makes it difficult to directly apply mainstream supervised learning models based on paired data to this scenario. Most existing image enhancement methods rely on training with low-light to normal-light paired datasets under ideal conditions. However, in practical applications of microscopy, obtaining large-scale, high-quality, and consistently labeled paired images is extremely costly and technically challenging. Therefore, researching an enhancement method that does not rely on paired training data and can simultaneously improve image contrast, brightness distribution, and detail sharpness while maintaining the structural integrity of label-free images has become an urgent need to drive technological development in this field. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a perception-driven enhancement network method and system for label-free microscopic cell images. By integrating techniques such as hybrid adaptive Retinex preprocessing, multi-network collaborative illumination-reflection decomposition, multi-scale noise perception, and dynamic Gamma correction, high-quality and structure-preserving enhancement of label-free microscopic cell images is achieved under unpaired data conditions.

[0006] On the one hand, perception-driven augmentation network methods for label-free microscopic cell images include:

[0007] S1. The acquired raw label-free microscopic cell image is input into the hybrid adaptive Retinex preprocessing module. Adaptive noise region detection is performed on the input image, dividing it into high and low noise regions and denoising them separately to obtain a denoised image. Multi-scale Retinex enhancement is applied to the denoised image to obtain an enhanced image. Edge compensation and S-curve stretching are performed on the enhanced image to obtain a detail-enhanced image. In the absence of paired data, the detail-enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image.

[0008] S2, using N-Net, L-Net and R-Net to estimate the illumination and reflectance components of label-free microscopic cell images, and obtain the illumination and reflectance components;

[0009] S3, input the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level;

[0010] S4 outputs a Gamma correction parameter that matches the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module;

[0011] S5, input the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determine the high and low noise regions based on the threshold, perform partitioned enhancement and correction, and output the noise sub-corrected image components.

[0012] S6. The noise-corrected image component and the reflection component are multiplied element-wise to obtain the noise-corrected image component. The noise-corrected image component and the auxiliary input image are then weighted and fused to obtain the final enhanced label-free microscopic cell image.

[0013] Furthermore, the method for obtaining the auxiliary input image is as follows:

[0014] First, calculate the local variance map, then generate the corresponding noise mask. The calculation formula is as follows:

[0015] ;

[0016] ;

[0017] in, Represents a point around which A local window; Indicates the region The number of pixels in; Indicates the window mean; Indicates the image at position Pixel values; Represents a constant factor; Indicates calculation in the region Variance of intra-pixel values; This represents a binary mask; This represents all local variance values ​​in a label-free microscopic cell image, i.e., all The median; based on the above two equations, distinguish between high-noise and low-noise regions;

[0018] Next, hybrid denoising is performed, applying NLM filtering to the high and low noise regions. The calculation formula is as follows:

[0019] ;

[0020] For low-noise areas, the original pixel values ​​are retained, calculated using the following formula:

[0021] ;

[0022] in, Indicates the location The pixel values ​​after noise reduction; Indicates The central local area; , which is a weighting coefficient used to determine the contribution of each pixel in the neighborhood to the center pixel; Indicates the image at position Pixel values;

[0023] Next, the denoised pixel values The multi-scale illumination components are calculated, the reflection components are extracted, and then fused. The calculation formula is as follows:

[0024] ;

[0025] ;

[0026] in, Represented at different scales The image under Gaussian blur; The standard deviation is expressed as Gaussian kernel; Indicates the convolution operation; Representing different scales; Gaussian kernel The scales are respectively =5, =15, =30; Indicates enhanced results; Used to avoid logarithms being negative infinity;

[0027] Finally, edge compensation and S-curve stretching are performed, calculated using the following formulas:

[0028] ;

[0029] ;

[0030] in, This represents the sharpened image; Indicates the sharpening intensity coefficient; This represents the final enhanced image. In order to control The slope of the function; The threshold parameter determines the enhancement center point; The reflection component of the sharpened image;

[0031] Label-free cell images The input image is processed using a hybrid adaptive Retinex module, resulting in an output image with optimized illumination and consistent structure. .

[0032] Furthermore, in S2, the label-free microscopic cell image is estimated using N-Net, L-Net, and R-Net to obtain the illumination and reflectance components. The calculation formulas are as follows: ; ; ; ; ; ;in, and These represent the original illumination component and the original reflection component, respectively; Represents the column coordinates of pixels in a label-free microscopic cell image; Represents the row coordinates of pixels in a label-free microscopic cell image; Is the image in position Pixel values; and Indicates the lighting component; and Indicates the reflected component; This represents a 3×3 convolution; and This represents the activation function.

[0033] Furthermore, in S3, the illumination component and reflection component are input into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level. The calculation formula is as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] in, The lighting components during the testing phase; Indicates low-frequency components; Indicates high-frequency components; Indicates the light sensitivity coefficient; Indicates the local noise level; Indicates low-frequency images in Gradient magnitude at; This represents the maximum gradient magnitude in the entire low-frequency image. It is an infinitesimal positive number that prevents division by zero; Indicates window size; This represents the local mean.

[0038] Furthermore, the calculation formula for S4 is as follows:

[0039] ;

[0040] ;

[0041] in, Indicates feature concatenation operation; This represents the entropy value of the reflection component, used to reflect the complexity of the region; Indicates the reflected component; This represents the concatenated input before it enters the lightweight CNN; Represents a linear transformation; Indicates the activation function; This represents the Gamma value of each pixel, which is predicted and output by a lightweight CNN.

[0042] Furthermore, the calculation formula for the noise sub-corrected image component is as follows:

[0043] ;

[0044] The correction formula for low-noise regions is as follows:

[0045]

[0046] The correction formula for high-noise regions is as follows:

[0047]

[0048] The final output noise-corrected image component:

[0049]

[0050] in, Indicates the adaptive noise threshold. This indicates the calculation of reflectance components in label-free microscopic cell images. Local entropy, Correction components for low-noise regions, Correction components for high-noise regions, For noise-corrected image components; This represents the mean local entropy of the reflectance component in a label-free microscopic cell image. ; This indicates that the luminance component is subjected to Gaussian filtering.

[0051] Furthermore, the final enhanced label-free microscopic cell image is calculated using the following formula:

[0052]

[0053]

[0054] in, This represents the enhanced luminance component; Represents the reflectance component of the original, unlabeled microscopic cell image; These are preliminary enhancement results; This is to enhance the results; For the final enhanced label-free microscopic cell image.

[0055] On the other hand, a perception-driven enhancement network system for label-free microscopic cell images includes:

[0056] The auxiliary input image acquisition module is used to input the acquired raw label-free microscopic cell image into the hybrid adaptive Retinex preprocessing module. The module performs adaptive noise region detection on the input image, divides it into high and low noise regions, and performs denoising processing on each region to obtain a denoised image. Multi-scale Retinex enhancement is applied to the denoised image to obtain an enhanced image. Edge compensation and S-curve stretching are then performed on the enhanced image to obtain a detail-enhanced image. In the absence of paired data, the detail-enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image.

[0057] The component acquisition module is used to estimate the illumination component and reflectance component of label-free microscopic cell images using N-Net, L-Net and R-Net, and to acquire the illumination component and reflectance component.

[0058] The decomposition module is used to input the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level;

[0059] The calibration parameter acquisition module is used to output Gamma calibration parameters that match the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module.

[0060] The correction module is used to input the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determine the high and low noise regions based on the threshold, perform partitioned enhancement and correction, and output the noise sub-corrected image component.

[0061] The weighted fusion module is used to multiply the noise sub-corrected image component and the reflection component element by element to obtain the noise corrected image component, and then weightedly fuse the noise corrected image component and the auxiliary input image to obtain the final enhanced label-free microscopic cell image.

[0062] The present invention adopts the above technical solution and has the following beneficial effects:

[0063] (1) The present invention performs noise detection, partition denoising and multi-scale enhancement on the original label-free microscopic cell image by using a hybrid adaptive Retinex preprocessing module, combined with edge compensation and S-curve stretching, to effectively improve the quality and detail of the image.

[0064] (2) This invention uses N-Net, L-Net and R-Net networks to jointly estimate the illumination component and reflection component of microscopic cell images, and extracts the illumination sensitivity coefficient and local noise level through a multi-scale illumination decomposition module to accurately separate the illumination component and reflection component in the image, providing accurate basic data support for subsequent Gamma correction, thereby improving the visual effect of the final output image.

[0065] (3) The present invention predicts pixel-level Gamma correction parameters based on local illumination and noise characteristics, and applies different enhancement strategies to regions with different noise levels in the adaptive noise detection enhancement module, so as to automatically adjust the brightness of each pixel according to the specific conditions of each pixel, ensuring that high-quality, high-contrast microscopic cell images can be generated even under high noise or low light conditions. Attached Figure Description

[0066] Figure 1 This is a flowchart of a perception-driven enhancement method for label-free microscopic cell images according to an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the overall framework of an embodiment of the present invention;

[0068] Figure 3 This is a schematic diagram of the hybrid adaptive Retinex module according to an embodiment of the present invention;

[0069] Figure 4 This is a schematic diagram illustrating the application of the present invention in the original and corrected images of unlabeled microscopic images according to an embodiment of the invention;

[0070] Figure 5 This is a diagram of a perception-driven enhancement system for label-free microscopic cell images according to an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0072] like Figure 1 As shown, the present invention is based on perception-driven enhancement for label-free microscopic cell images, including:

[0073] S1. The acquired raw label-free microscopic cell image is input into the hybrid adaptive Retinex preprocessing module. Adaptive noise region detection is performed on the input image, dividing it into high and low noise regions and denoising them separately to obtain a denoised image. Multi-scale Retinex enhancement is applied to the denoised image to obtain an enhanced image. Edge compensation and S-curve stretching are performed on the enhanced image to obtain a detail-enhanced image. In the absence of paired data, the detail-enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image.

[0074] Specifically, the method for obtaining the auxiliary input image is as follows:

[0075] First, calculate the local variance map, then generate the corresponding noise mask. The calculation formula is as follows:

[0076] ;

[0077] ;

[0078] in, Represents a point around which A 15x15 local window; Indicates the region The number of pixels in; Indicates the window mean; Indicates the image at position Pixel values; Represents a constant factor; Indicates calculation in the region Variance of intra-pixel values; This represents a binary mask; This represents all local variance values ​​in a label-free microscopic cell image, i.e., all The median; based on the above two equations, distinguish between high-noise and low-noise regions;

[0079] Next, hybrid denoising is performed, applying NLM filtering to the high and low noise regions. The calculation formula is as follows:

[0080] ;

[0081] For low-noise areas, the original pixel values ​​are retained, calculated using the following formula:

[0082] ;

[0083] in, Indicates the location The pixel values ​​after noise reduction; Indicates The central local area; , which is a weighting coefficient used to determine the contribution of each pixel in the neighborhood to the center pixel; Indicates the image at position Pixel values;

[0084] Next, the denoised pixel values The multi-scale illumination components are calculated, the reflection components are extracted, and then fused. The calculation formula is as follows:

[0085] ;

[0086] ;

[0087] in, Represented at different scales The image under Gaussian blur; The standard deviation is expressed as Gaussian kernel; Indicates the convolution operation; Representing different scales; Gaussian kernel The scales are respectively =5, =15, =30; Indicates enhanced results; Used to avoid logarithms being negative infinity;

[0088] Finally, edge compensation and S-curve stretching are performed, calculated using the following formulas:

[0089] ;

[0090] ;

[0091] in, This represents the sharpened image; Indicates the sharpening intensity coefficient; This represents the final enhanced image. In order to control The slope of the function. The threshold parameter determines the enhancement center point. The reflection component of the sharpened image;

[0092] Label-free cell images The input image is processed using a hybrid adaptive Retinex module, resulting in an output image with optimized illumination and consistent structure. .

[0093] Specifically, such as Figure 2This is a schematic diagram of the overall framework of an embodiment of the present invention. To address the problem of the lack of standard paired data in the field of label-free microscopic cell images, the present invention designs a hybrid adaptive Retinex module. During the training phase, the acquired original label-free microscopic cell images are input into the hybrid adaptive Retinex preprocessing module. In the absence of paired data, the image with enhanced details is output as an auxiliary image that maintains structural consistency with the input image, thereby obtaining an auxiliary input image. During the testing phase, N-Net, L-Net, and R-Net were used to estimate the illumination and reflectance components of label-free microscopic cell images, obtaining these components. The illumination and reflectance components were then input into a multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level. A dynamic Gamma parameter prediction module then outputs a Gamma correction parameter matching the illumination sensitivity coefficient and local noise level. The matched Gamma correction parameter and illumination components were input into an adaptive noise detection and enhancement module. High and low noise regions were identified based on thresholds, and partitioned enhancement and correction were performed, outputting a noise-corrected image component. The noise-corrected image component and reflectance component were multiplied element-wise to obtain the noise-corrected image component. Finally, the noise-corrected image component and the auxiliary input image were weighted and fused to obtain the final enhanced label-free microscopic cell image, achieving the effect of enhancing the label-free microscopic cell image. Figure 3 The diagram shows a schematic of the hybrid adaptive Retinex module in an embodiment of the present invention. The acquired original label-free microscopic cell image is input into the hybrid adaptive Retinex preprocessing module. The input image is subjected to adaptive noise region detection, which divides it into high and low noise regions and performs denoising processing on them respectively to obtain a denoised image. Multi-scale Retinex enhancement is applied to the denoised image to obtain an enhanced image. Edge compensation and S-curve stretching are performed on the enhanced image to obtain a detail-enhanced image. In the absence of paired data, the detail-enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image.

[0094] S2 uses N-Net, L-Net, and R-Net to estimate the illumination and reflection components of label-free microscopic cell images, thus obtaining the illumination and reflection components.

[0095] Specifically, the illumination and reflectance components of label-free microscopic cell images are estimated using N-Net, L-Net, and R-Net, and the calculation formulas are as follows: ; ; ; ; ; ;in, and These represent the original illumination component and the original reflection component, respectively; Represents the column coordinates of pixels in a label-free microscopic cell image; Represents the row coordinates of pixels in a label-free microscopic cell image; Is the image in position Pixel values; and Indicates the lighting component; and Indicates the reflected component; This represents a 3×3 convolution; and This represents the activation function.

[0096] Specifically, in this embodiment, L-Net and R-Net are used to estimate the illumination component and reflectance component, respectively. These two network structures are very similar and simple in design, each consisting of five convolutional layers. The first four convolutional layers of L-Net use ReLU as the activation function, while the last layer uses the Sigmoid function to normalize the output to the range [0,1]. The first five convolutional layers of R-Net use ReLU as the activation function. Based on Retinex theory, we assume that the three color channels share the same illumination information. Experimental results show that even with such a simple network structure, relatively ideal performance has been achieved. In addition to L-Net and R-Net, we also introduce an N-Net, which is structurally similar to L-Net, to suppress inappropriate features in the original image.

[0097] S3 inputs the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level.

[0098] Specifically, the illumination component and reflection component are input into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level. The calculation formula is as follows:

[0099] ;

[0100] ;

[0101] ;

[0102] in, The lighting components during the testing phase; Indicates low-frequency components; Indicates high-frequency components; Indicates the light sensitivity coefficient; Indicates the local noise level; Indicates low-frequency images in Gradient magnitude at; This represents the maximum gradient magnitude in the entire low-frequency image. It is an infinitesimal positive number that prevents division by zero; Indicates window size; This represents the local mean.

[0103] Specifically, the calculation of the light sensitivity coefficient pays special attention to cell biological characteristics—automatically reducing the enhancement intensity in densely celled regions (such as cell junctions) to avoid membrane structure fusion, while enhancing the signal intensity in low-contrast regions within the cell (such as mitochondrial distribution areas). Simultaneously, based on a dynamic adjustment mechanism of local noise levels, conservative denoising is employed in the optical diffraction-limited region (≈200nm) of microscopic imaging to preserve the true structure, while strong noise suppression is implemented in non-focal plane regions.

[0104] S4 outputs a Gamma correction parameter that matches the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module.

[0105] Specifically, to achieve pixel-level adaptive enhancement tailored to the characteristics of cell images, this embodiment designs a dynamic Gamma prediction module that generates spatially varying Gamma correction maps in real time using a lightweight CNN network. Based on the reflectance component of the label-free microscopic cell image and the illumination sensitivity coefficient and local noise level obtained from the multi-scale illumination decomposition module, this module dynamically adjusts the Gamma value of each pixel. The module innovatively introduces a noise perception mechanism, automatically constraining the enhancement amplitude based on the local signal-to-noise ratio, limiting the enhancement factor to within 1.2 times in high-noise regions, effectively suppressing the noise amplification problem common in traditional methods. The specific process is as follows: first, an input feature map is constructed, and three-channel features are concatenated as input to the parameter prediction subnet.

[0106] Specifically, the calculation formula for S4 is as follows:

[0107] ;

[0108] ;

[0109] in, Indicates feature concatenation operation; This represents the entropy value of the reflection component, used to reflect the complexity of the region; Indicates the reflected component; This represents the concatenated input before it enters the lightweight CNN; Represents a linear transformation; Indicates the activation function; This represents the Gamma value of each pixel, which is predicted and output by a lightweight CNN.

[0110] S5 inputs the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determines the high and low noise regions based on the threshold, performs partitioned enhancement and correction, and outputs the noise sub-corrected image components.

[0111] Specifically, in S5, the calculation formula for the noise sub-corrected image component is as follows:

[0112] ;

[0113] The correction formula for low-noise regions is as follows:

[0114]

[0115] The correction formula for high-noise regions is as follows:

[0116]

[0117] The final output noise-corrected image component:

[0118]

[0119] in, Indicates the adaptive noise threshold. This indicates the calculation of reflectance components in label-free microscopic cell images. Local entropy, Correction components for low-noise regions, Correction components for high-noise regions, For noise-corrected image components; This represents the mean local entropy of the reflectance component in a label-free microscopic cell image. ; This indicates that the luminance component is subjected to Gaussian filtering.

[0120] Specifically, the Tnoise of a given unlabeled microscopic cell image is calculated; if it is less than Tnoise, it is considered a low-noise region, and if it is greater than Tnoise, it is considered a high-noise region.

[0121] S6. The noise-corrected image component and the reflection component are multiplied element-wise to obtain the noise-corrected image component. The noise-corrected image component and the auxiliary input image are then weighted and fused to obtain the final enhanced label-free microscopic cell image.

[0122] Specifically, the formula for calculating the final enhanced label-free microscopic cell image is as follows:

[0123]

[0124]

[0125] in, This represents the enhanced luminance component; Represents the reflectance component of the original, unlabeled microscopic cell image; These are preliminary enhancement results; This is to enhance the results; For the final enhanced label-free microscopic cell image.

[0126] Specifically, in this embodiment, to ensure input quality, unsuitable features are preferentially removed, rather than performing deep Retinex decomposition on the original low-light image. To this end, we constructed a dedicated loss function to guide this process, calculated as follows:

[0127]

[0128] in The input image represents the raw, unlabeled microscopic image of cells to be processed. Represents the projected image. Denotes the square of the L2 norm. This represents the projection loss. Loss function. The original image is converted into a specific image that is more suitable for Retinex decomposition.

[0129] Specifically, since the noise component in the Retinex model is not considered, the image is expected to be noise-free. In addition, some useless features are discarded at this stage. The projection loss can be explained from the perspective of error redistribution, and the calculation formula is as follows:

[0130] ;

[0131] in, This indicates the search for the parameter variable that minimizes the objective function. This represents an intermediate image, optimized using the projection loss function, that is more suitable for Retinex decomposition. This represents the input label-free microscopic cell image. This represents the illumination component to be solved. This represents the reflection component to be solved. This represents the noise term in the model. Describing the Frobenius norm, This represents element-wise multiplication. The goal is to shift the error from Retinex decomposition to the projection stage. Since the Retinex model cannot represent sensor noise, the projected image is highly similar to the original image. It's important to note that the projection loss needs to be combined with other constraints to avoid trivial solutions.

[0132] Reflection consistency loss It is calculated based on low-light image pairs and Retinex theory. This is compared to prior knowledge of handcrafted methods. It is more accurate and adaptive because it reveals the physical properties of an object. Mathematically, It is expressed as:

[0133] ;

[0134] in, As the normalization factor, This represents the total number of pixels. and For the reflection components of a pair of low-light images, This indicates that the network is forced to predict the same reflectance component. When contrast is increased, noise hidden in dark areas is amplified. To address this noise issue, existing methods either add a smoothing term to the estimated reflectance or perform denoising after enhancement. Since the network utilizes Retinex theory to decompose low-light images into illuminance and reflectance components, some basic constraints on Retinex decomposition are introduced. The Retinex loss can be expressed as:

[0135] ;

[0136] in, This represents an intermediate image, optimized using the projection loss function, that is more suitable for Retinex decomposition. This represents the initial estimate of illuminance. Representing the horizontal and vertical gradients. First, the decomposed components should meet the requirements for reconstructing the input image. Therefore, a reconstruction term is used. To ensure reasonable decomposition. Once the illuminance is estimated, reflectance can be calculated by pixel-by-pixel partitioning between the label-free microscopic cell image and its illuminance map. Specifically, initial illumination... It is calculated using the maximum values ​​of the R, G, and B channels, and the calculation formula is as follows:

[0137] ;

[0138] in, The index variable represents the color channels, namely the red channel, green channel, and blue channel. Indicates the pixel position Color channel The intensity value.

[0139] Specifically, the overall loss function for training the network is a linear combination of each loss, calculated as follows:

[0140] ;

[0141] in, , and Indicates the weight.

[0142] In summary, the advantages of this invention are as follows: Addressing the lack of standard paired data in the field of label-free microscopic cell images, and the common problems of low contrast, insufficient brightness, and blurred cell edge details in such images, this invention proposes a microscopic image enhancement network based on adaptive Retinex theory. First, to address the problem of missing paired data, a hybrid adaptive Retinex preprocessing module is designed. Through noise-aware multi-scale enhancement and edge compensation mechanisms, it generates auxiliary input images with consistent structure without requiring paired data. Second, considering the unique biological structural features of label-free microscopic images, a multi-scale illumination decomposition module is constructed, which can effectively separate the low-frequency illumination component containing the overall cell morphology and the high-frequency detail component carrying subcellular structural information. Furthermore, to achieve pixel-level adaptive enhancement, a dynamic Gamma prediction module is designed, utilizing a lightweight CNN to generate spatially adaptive Gamma correction maps in real time. This module dynamically adjusts the Gamma value of each pixel based on the image's reflectance component, illumination sensitivity coefficient, and local noise level. It innovatively introduces a noise perception mechanism, automatically limiting the enhancement intensity based on the local signal-to-noise ratio—controlling the enhancement factor to within 1.2 times in high-noise regions, thus significantly suppressing the noise amplification phenomenon common in traditional enhancement methods. Finally, an adaptive noise detection enhancement module is designed, dynamically selecting the enhancement strategy based on the entropy value of the reflectance component. A dynamic threshold is used to distinguish between noise and effective regions, effectively suppressing noise while significantly improving image details and edge features.

[0143] Specifically, such as Figure 4 The figure shows an example of the application of the present invention to the original and corrected images of a label-free microscopic image. As can be seen from the figure, the present invention has an excellent effect on enhancing label-free microscopic images, enhancing the characterization of edge details and minute features.

[0144] Specifically, in this embodiment, the acquired label-free microscopic cell image is input to the hybrid adaptive Retinex preprocessing module and N-Net, L-Net, and R-Net to obtain the original auxiliary input image, illumination component, and reflectance component, respectively. The N-Net and L-Net each consist of 5 3×3 convolutional layers, 4 ReLU activation layers, and 1 Sigmoid activation layer, while the R-Net consists of 5 3×3 convolutional layers and 5 ReLU activation layers. The illumination component is decomposed into low-frequency and high-frequency components, and the illumination sensitivity coefficient and local noise level are calculated. The reflectance component, illumination sensitivity coefficient, and local noise level are input to the lightweight weighted array. The matching Gamma correction parameters are calculated using a heavy CNN convolution; the lightweight weighted CNN convolution consists of two 3×3 convolutional layers, two ReLU activation layers, one FCLayer, and one Sigmoid activation layer; the enhancement strategy is dynamically selected based on the entropy value of the reflection component and the local noise level, and the illumination component is dynamically enhanced in different regions, outputting noise sub-corrected image components after Gamma correction in different regions; the noise sub-corrected image component is multiplied element-wise with the reflection component to obtain the noise-corrected image component; then it is fused with the label-free microscopic cell auxiliary input image processed by hybrid adaptive Retinex to output the final enhanced label-free microscopic cell image.

[0145] like Figure 5 As shown, this embodiment also discloses a perception-driven enhancement system based on label-free microscopic cell images, including:

[0146] The auxiliary input image acquisition module 51 is used to input the acquired raw label-free microscopic cell image into the hybrid adaptive Retinex preprocessing module, perform adaptive noise region detection on the input image, divide it into high and low noise regions and perform denoising processing on them respectively to obtain a denoised image; apply multi-scale Retinex enhancement to the denoised image to obtain an enhanced image; perform edge compensation and S-curve stretching on the enhanced image to obtain a detail-enhanced image; in the absence of paired data, output the detail-enhanced image as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image;

[0147] The component acquisition module 52 is used to estimate the illumination component and reflectance component of the label-free microscopic cell image through N-Net, L-Net and R-Net, and to acquire the illumination component and reflectance component.

[0148] Decomposition module 53 is used to input the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level;

[0149] The correction parameter acquisition module 54 is used to output Gamma correction parameters that match the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module.

[0150] The correction module 55 is used to input the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determine the high and low noise regions based on the threshold, perform partitioned enhancement and correction, and output the noise sub-corrected image components.

[0151] The weighted fusion module 56 is used to multiply the noise sub-corrected image component and the reflection component element by element to obtain the noise corrected image component, and to weightedly fuse the noise corrected image component and the auxiliary input image to obtain the final enhanced label-free microscopic cell image.

[0152] The specific implementation of the perception-driven enhancement network system for label-free microscopic cell images is the same as that of the perception-driven enhancement network method for label-free microscopic cell images, and will not be described again in this embodiment.

[0153] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A perception driven augmentation network method for markerless microscopic cell images, characterized in that, Includes the following steps: S1, the acquired raw label-free microscopic cell image is input into the hybrid adaptive Retinex preprocessing module, which performs adaptive noise region detection on the input image, divides it into high and low noise regions, and performs denoising processing on them respectively to obtain the denoised image; Multi-scale Retinex enhancement is applied to the denoised image to obtain the enhanced image. Edge compensation and S-curve stretching are then performed on the enhanced image to obtain the detailed enhanced image. In the absence of paired data, the detailed enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image. S2, using N-Net, L-Net and R-Net to estimate the illumination and reflectance components of label-free microscopic cell images, and obtain the illumination and reflectance components; S3, input the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level; S4 outputs a Gamma correction parameter that matches the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module; S5, input the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determine the high and low noise regions based on the threshold, perform partitioned enhancement and correction, and output the noise sub-corrected image components. S6, multiply the noise sub-corrected image component and the reflection component element by element to obtain the noise corrected image component, and then weight and fuse the noise corrected image component and the auxiliary input image to obtain the final enhanced label-free microscopic cell image; The formula for calculating S4 is as follows: ； ； in, Indicates feature concatenation operation; Indicates the light sensitivity coefficient; Indicates the local noise level; This represents the entropy value of the reflection component, used to reflect the complexity of the region; Indicates the reflected component; This represents the concatenated input before it enters the lightweight CNN; Represents a linear transformation; and Both represent activation functions; This represents the Gamma value of each pixel, which is predicted and output by a lightweight CNN. The calculation formula for the noise sub-corrected image component is as follows: ； The correction formula for low-noise regions is as follows: ； The correction formula for high-noise regions is as follows: The final output noise-corrected image component is: ； in, Indicates the adaptive noise threshold. This indicates the calculation of reflectance components in label-free microscopic cell images. Local entropy, Correction components for low-noise regions, Correction components for high-noise regions, For noise-corrected image components; This represents the mean local entropy of the reflectance component in a label-free microscopic cell image. ; This indicates that the luminance component is subjected to Gaussian filtering; This refers to the lighting components used during the testing phase.

2. The perception-driven enhancement network method for label-free microscopic cell images according to claim 1, characterized in that, In S1, the method for obtaining the auxiliary input image is as follows: First, calculate the local variance map, then generate the corresponding noise mask. The calculation formula is as follows: ； ； in, Represents a point around which A local window; Indicates the region The number of pixels in; Indicates the window mean; Indicates the image at position Pixel values; Represents a constant factor; Indicates calculation in the region Variance of intra-pixel values; This represents a binary mask; This represents all local variance values ​​in a label-free microscopic cell image, i.e., all The median; based on the above two equations, distinguish between high-noise and low-noise regions; Next, hybrid denoising is performed, applying NLM filtering to the high and low noise regions. The calculation formula is as follows: ； For low-noise areas, the original pixel values ​​are retained, calculated using the following formula: ； in, Indicates the location The pixel values ​​after noise reduction; Indicates The central local area; , which is a weighting coefficient used to determine the contribution of each pixel in the neighborhood to the center pixel; Indicates the image at position Pixel values; Next, the denoised pixel values The multi-scale illumination components are calculated, the reflection components are extracted, and then fused. The calculation formula is as follows: ； ； in, Represented at different scales The image under Gaussian blur; The standard deviation is expressed as Gaussian kernel; Indicates the convolution operation; Representing different scales; Gaussian kernel The scales are respectively =5, =15, =30; Indicates enhanced results; Used to avoid logarithms being negative infinity; Finally, edge compensation and S-curve stretching are performed, calculated using the following formulas: ； ； in, This represents the sharpened image; Indicates the sharpening intensity coefficient; This represents the final enhanced image. In order to control The slope of the function; The threshold parameter determines the enhancement center point; The reflection component of the sharpened image; Label-free cell images The input image is processed using a hybrid adaptive Retinex module, resulting in an output image with optimized illumination and consistent structure. .

3. The perception-driven enhancement network method for label-free microscopic cell images according to claim 1, characterized in that, In S2, label-free microscopic cell images are estimated using N-Net, L-Net, and R-Net to obtain the illumination and reflectance components. The calculation formulas are as follows: ; ; ; ; ; ;in, and These represent the original illumination component and the original reflection component, respectively; Represents the column coordinates of pixels in a label-free microscopic cell image; Represents the row coordinates of pixels in a label-free microscopic cell image; Is the image in position Pixel values; and Indicates the lighting component; and Indicates the reflected component; This represents a 3×3 convolution; and This represents the activation function.

4. The perception-driven enhancement network method for label-free microscopic cell images according to claim 1, characterized in that, In S3, the illumination component and reflection component are input into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level. The calculation formula is as follows: ； ； ； in, The lighting components during the testing phase; Indicates low-frequency components; Indicates high-frequency components; Indicates the light sensitivity coefficient; Indicates the local noise level; Indicates low-frequency images in Gradient magnitude at; This represents the maximum gradient magnitude in the entire low-frequency image. It is an infinitesimal positive number that prevents division by zero; Represents a point around which A local window; This represents the window mean.

5. The perception-driven enhancement network method for label-free microscopic cell images according to claim 1, characterized in that, In S6, the formula for calculating the final enhanced label-free microscopic cell image is as follows: ； ； in, This represents the enhanced luminance component; Represents the reflectance component of the original, unlabeled microscopic cell image; These are preliminary enhancement results; This is to enhance the results; For the final enhanced label-free microscopic cell image.

6. A perception-driven enhancement network system based on label-free microscopic cell images, characterized in that, include: The auxiliary input image acquisition module is used to input the acquired raw label-free microscopic cell image into the hybrid adaptive Retinex preprocessing module, which performs adaptive noise region detection on the input image, divides it into high and low noise regions, and performs denoising processing on them respectively to obtain the denoised image. Multi-scale Retinex enhancement is applied to the denoised image to obtain the enhanced image. Edge compensation and S-curve stretching are then performed on the enhanced image to obtain the detailed enhanced image. In the absence of paired data, the detailed enhanced image is output as an auxiliary image that maintains structural consistency with the input image to obtain an auxiliary input image. The component acquisition module is used to estimate the illumination component and reflectance component of label-free microscopic cell images using N-Net, L-Net and R-Net, and to acquire the illumination component and reflectance component. The decomposition module is used to input the illumination component and reflection component into the multi-scale illumination decomposition module to obtain the illumination sensitivity coefficient and local noise level; The calibration parameter acquisition module is used to output Gamma calibration parameters that match the light sensitivity coefficient and local noise level through the dynamic Gamma parameter prediction module. The correction module is used to input the matched Gamma correction parameters and illumination components into the adaptive noise detection and enhancement module, determine the high and low noise regions based on the threshold, perform partitioned enhancement and correction, and output the noise sub-corrected image component. The weighted fusion module is used to multiply the noise sub-corrected image component and the reflection component element by element to obtain the noise corrected image component, and then weightedly fuse the noise corrected image component and the auxiliary input image to obtain the final enhanced label-free microscopic cell image. The calculation formula for the Gamma correction parameter, which is matched with the illumination sensitivity coefficient and local noise level, output by the dynamic Gamma parameter prediction module is as follows: ； ； in, Indicates feature concatenation operation; Indicates the light sensitivity coefficient; Indicates the local noise level; This represents the entropy value of the reflection component, used to reflect the complexity of the region; Indicates the reflected component; This represents the concatenated input before it enters the lightweight CNN; Represents a linear transformation; and Both represent activation functions; This represents the Gamma value of each pixel, which is predicted and output by a lightweight CNN. The calculation formula for the noise sub-corrected image component is as follows: ； The correction formula for low-noise regions is as follows: ； The correction formula for high-noise regions is as follows: The final output noise-corrected image component is: ； in, Indicates the adaptive noise threshold. This indicates the calculation of reflectance components in label-free microscopic cell images. Local entropy, Correction components for low-noise regions, Correction components for high-noise regions, For noise-corrected image components; This represents the mean local entropy of the reflectance component in a label-free microscopic cell image. ; This indicates that the luminance component is subjected to Gaussian filtering; This refers to the lighting components used during the testing phase.

Images