Explainable farmland image enhancement method and system based on physical perception and reinforcement learning

Abstract

The present application relates to the technical field of image processing, in particular to an interpretable farmland image enhancement method based on physical perception and reinforcement learning, comprising: obtaining an original farmland image and performing global perception analysis, extracting multi-scale features and identifying degraded areas and feature distribution; based on the global perception analysis result, constructing a semantic enhancement blueprint of quantified physical properties of degraded areas and optimization requirements; initializing a reinforcement learning agent according to the semantic enhancement blueprint, selecting an image processing operation sequence through a physical constraint reward function; executing the image operation sequence, terminating the enhancement processing according to the quality evaluation index increment threshold and the physical consistency of the semantic enhancement blueprint; outputting an enhanced image and an interpretable report of the physical basis traceability of the associated image operation sequence. The purpose is to solve the technical problems of lack of interpretability, insufficient environmental adaptability and rigid decision mechanism of farmland image enhancement technology.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an interpretable farmland image enhancement method and system based on physical perception and reinforcement learning. Background Technology

[0002] Farmland image processing technology, as a key support for the development of precision agriculture and smart agriculture, aims to achieve important functions such as crop growth monitoring, early identification of pests and diseases, and dynamic yield prediction through intelligent analysis of collected farmland images. However, the complexity and dynamism of farmland scenes pose significant challenges to acquiring high-quality images. During field image acquisition, extreme lighting conditions such as strong direct sunlight, staggered shadows, and backlighting can easily lead to local overexposure or underexposure, causing overall contrast imbalance. Atmospheric interference such as haze, dust, and water vapor can significantly reduce image contrast, causing color distortion and blurring key crop details. At the same time, equipment vibration and flight altitude fluctuations during drone aerial photography, as well as environmental changes brought about by crop growth cycles and seasonal changes, further exacerbate the instability of image quality and the difficulty of standardized processing. These common image quality problems, such as blurred details, color deviation, and uneven lighting, directly restrict the accuracy and reliability of subsequent automated analysis, becoming a major bottleneck restricting the progress of agricultural intelligence.

[0003] To address these challenges, existing farmland image enhancement technologies mainly fall into three categories, but their limitations make them insufficient for practical applications. Traditional image processing methods, such as histogram equalization, contrast stretching, and filtering denoising, are simple and computationally efficient, but they are poorly adaptable to the complex and variable lighting and environmental conditions of farmland, often resulting in over-enhancement leading to noise amplification or under-enhancement causing detail loss. Deep learning-based methods, especially end-to-end models built from convolutional neural networks and generative adversarial networks, show certain advantages in specific scenarios, but their "black box" nature leads to a lack of physical interpretability in the enhancement process, raising questions about the reliability and traceability of the results. Furthermore, they heavily rely on large-scale, accurately labeled data, which is extremely costly to obtain in farmland scenarios. Physical model-based methods attempt to simulate degradation processes through models such as atmospheric scattering and light attenuation. While theoretically sound, they struggle to accurately estimate model parameters in complex farmland environments, resulting in limited generalization ability and unstable processing effects. The methods mentioned above together form the technical foundation for current agricultural image enhancement. However, existing methods still generally suffer from problems such as lack of interpretability, insufficient environmental adaptability, rigid decision-making mechanisms, and lack of physical constraints, making it difficult to meet the requirements of agricultural applications for processing transparency, scene robustness, and result reliability. Summary of the Invention

[0004] To address the technical problems of lack of interpretability, insufficient environmental adaptability, and rigid decision-making mechanisms in farmland image enhancement technology, and to achieve a complete enhancement process from physical perception and dynamic decision-making to interpretable output, this invention provides an interpretable farmland image enhancement method and system based on physical perception and reinforcement learning. The specific technical solution adopted is as follows:

[0005] The first aspect of the present invention provides an interpretable farmland image enhancement method based on physical perception and reinforcement learning, the method comprising:

[0006] Acquire raw farmland images and perform global perception analysis to extract multi-scale features and identify degraded areas and feature distributions;

[0007] Based on the global perception analysis results, a semantic enhancement blueprint is constructed to quantify the physical attributes and optimization requirements of the degraded region;

[0008] The reinforcement learning agent is initialized according to the semantic enhancement blueprint, and the image processing operation sequence is selected through the physical constraint reward function;

[0009] The image processing operation sequence is executed, and the enhancement process is terminated based on the incremental threshold of the quality assessment index and the physical consistency of the semantic enhancement blueprint.

[0010] Output an interpretability report that provides physical evidence for the enhanced image and associated image processing sequence.

[0011] Furthermore, original farmland images are acquired and global perception analysis is performed to extract multi-scale features and identify degraded areas and feature distributions, including:

[0012] The original farmland image was size-normalized.

[0013] Convert the RGB color space to the LAB color space to separate luminance and chrominance information;

[0014] A U-shaped Transformer network embedded with multi-dimensional physical prior knowledge is used to extract multi-scale features and identify degraded regions and feature distributions. The multi-dimensional physical prior knowledge includes atmospheric scattering prior, vegetation spectrum prior, illumination model prior, and depth distance prior.

[0015] Furthermore, a semantically enhanced blueprint is constructed to quantify the physical attributes of the degraded region and the optimization requirements, including:

[0016] Based on the physical model of light attenuation, the transmittance distribution is calculated, and a spatial attenuation feature map of the region with blurred markings and color deviation is generated.

[0017] The degree of channel degradation is determined by analyzing the variance distribution of each color channel, and a channel priority feature vector is generated.

[0018] Calculate the color difference between the current image and a standard farmland in the LAB color space to generate color correction coefficients;

[0019] The spatial attenuation feature map, channel priority vector, and color correction coefficients are combined and fused with vegetation mask estimation results to form the semantic enhancement blueprint.

[0020] Furthermore, based on the physical model of light attenuation, the transmittance distribution is calculated, and a spatial attenuation feature map of the blurred and color-shifted regions is generated, including:

[0021] A light attenuation model was established based on atmospheric scattering theory.

[0022] Estimating scene depth distribution using dark channel priors;

[0023] Transmittance distribution is calculated by integrating the atmospheric attenuation coefficient of farmland with scene depth distribution;

[0024] A spatial attenuation feature map is generated to identify highly blurred areas based on the transmittance threshold.

[0025] Furthermore, the color difference between the current image and a standard farmland is calculated in the LAB color space to generate color correction coefficients, including:

[0026] Obtain the average luminance and chrominance values ​​of the LAB channel of the current image;

[0027] Calculate the color offset of each channel based on the pre-stored standard values ​​for healthy farmland;

[0028] Color correction coefficients are generated based on the color offset of each channel for brightness correction and chromaticity compensation.

[0029] Further, initializing the reinforcement learning agent according to the semantic enhancement blueprint includes:

[0030] The initial state space of the reinforcement learning agent is configured based on the spatial decay feature map, channel priority feature vector and color correction coefficient of the semantic enhancement blueprint.

[0031] Configure an extended action set that includes physical boot operations and image processing operations;

[0032] A state-action value function is constructed based on the Q-learning algorithm, where the state representation includes the fusion features of the current image features and semantically enhanced blueprint information.

[0033] Furthermore, initializing the reinforcement learning agent based on the semantic enhancement blueprint also includes:

[0034] The processing priority order of each color channel is determined based on the channel priority feature vector;

[0035] The processing intensity parameters for different regions are set based on the spatial distribution characteristics of the spatial attenuation feature map.

[0036] The correction intensity parameters for each color channel are set based on the color correction coefficient.

[0037] Furthermore, the image processing operation sequence is selected through a physically constrained reward function, including:

[0038] The incremental value of the spatial fraction of the spatial decay feature map relative to the previous operation is calculated as the result of the operation.

[0039] Calculate the match between the current operation and the semantic enhancement blueprint, including:

[0040] The computational results are consistent with the spatial distribution of the spatial attenuation feature map.

[0041] The computational operation effect is consistent with the channel processing order of the channel priority feature vector;

[0042] The calculated operation effect is consistent with the color correction effect of the color correction coefficient;

[0043] By combining quality increments and matching degrees through weighted fusion, a reward value is generated, which drives the reinforcement learning agent to select the optimal action.

[0044] Furthermore, the enhancement process is terminated based on the incremental threshold of the quality assessment metrics and the physical consistency of the semantic enhancement blueprint, including:

[0045] Calculate the mean increment of the image quality assessment index after multiple operations;

[0046] Calculate the physical consistency alignment between the current operation sequence and the semantic enhancement blueprint, the alignment being calculated based on the consistency between the current operation sequence and the spatial decay feature map, the channel priority feature vector, and the color correction coefficient;

[0047] The process terminates when the mean increment is lower than the first threshold and the alignment is higher than the second threshold.

[0048] The second aspect of the present invention provides an interpretable farmland image enhancement system based on physical perception and reinforcement learning, employing the interpretable farmland image enhancement method based on physical perception and reinforcement learning described in the first aspect of the present invention. The system includes:

[0049] The image perception module is configured to acquire raw farmland images and perform global perception analysis, extract multi-scale features, and identify degraded areas and feature distributions.

[0050] The blueprint building module is configured to build a semantically enhanced blueprint that quantifies the physical attributes and optimization requirements of the degraded region based on the global perception analysis results.

[0051] The decision agent module is configured to initialize a reinforcement learning agent based on the semantic enhancement blueprint and select an image processing operation sequence through a physical constraint reward function;

[0052] The enhancement execution module is configured to execute the image processing operation sequence and terminate the enhancement processing based on the incremental threshold of the quality assessment index and the physical consistency of the semantic enhancement blueprint.

[0053] The output module is configured to output an interpretability report that traces the physical basis of the enhanced image and associated image processing operation sequence.

[0054] The present invention has the following beneficial effects:

[0055] This invention provides an interpretable farmland image enhancement method based on physical perception and reinforcement learning. By performing global perception analysis on the original farmland image, it extracts multi-scale features and identifies degraded regions and feature distributions. This accurately captures the spatial and feature dimension information of image degradation, avoiding the blind enhancement problem caused by the lack of identification of degradation distributions in existing methods. Based on the global perception results, a semantic enhancement blueprint is constructed to quantify the physical attributes and optimization requirements of degraded regions, directly linking the enhancement target with the physical degradation mechanism. This solves the deficiency of lacking physical constraints in existing purely data-driven methods, providing a deterministic physical basis and optimization direction for subsequent dynamic decision-making. Based on the semantic enhancement blueprint... This method initializes a reinforcement learning agent and selects image processing operation sequences using a physically constrained reward function. It achieves dynamic decision-making based on specific image degradation characteristics, improving adaptability to different farmland image degradation scenarios and addressing the rigidity of decision-making mechanisms in existing methods. Simultaneously, by using incremental thresholds for quality assessment metrics and physical consistency of the semantic enhancement blueprint as termination conditions during operation sequence execution, it ensures image enhancement effectiveness while avoiding over-enhancement or violations of physical laws, thus guaranteeing the reliability of the enhancement results. Finally, when outputting the enhanced image, it simultaneously provides an interpretable report tracing the physical basis of the associated operation sequences, achieving full transparency and traceability of the enhancement process. This method effectively overcomes the problems of traditional methods lacking physical basis, the black-box nature of deep learning, and the difficulty in estimating physical model parameters. It improves the transparency, adaptability, and reliability of the image enhancement process, providing more reliable technical support for the intelligent development of precision agriculture. Attached Figure Description

[0056] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of an interpretable farmland image enhancement method based on physical perception and reinforcement learning, provided in one embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the overall architecture of an interpretable farmland image enhancement method based on physical perception and reinforcement learning provided in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of a primitive farmland image provided in one embodiment of the present invention;

[0060] Figure 4 This is an enhanced image schematic diagram provided according to an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the structure of an interpretable farmland image enhancement system based on physical perception and reinforcement learning, provided in an embodiment of the present invention. Detailed Implementation

[0062] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an interpretable farmland image enhancement method and system based on physical perception and reinforcement learning proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0064] The following description, in conjunction with the accompanying drawings, details the specific scheme of the interpretable farmland image enhancement method and system based on physical perception and reinforcement learning provided by this invention.

[0065] To address the shortcomings of existing farmland image enhancement technologies, such as lack of interpretability, insufficient adaptability, and simplistic decision-making mechanisms, please refer to [link / reference needed]. Figure 1 The diagram illustrates a flowchart of an interpretable farmland image enhancement method based on physical perception and reinforcement learning, according to an embodiment of the present invention. The method includes:

[0066] Step S100: Acquire the original farmland image and perform global perception analysis to extract multi-scale features and identify degraded areas and feature distribution;

[0067] Step S100 specifically includes:

[0068] Step S110: Perform size normalization processing on the original farmland image;

[0069] In some embodiments, low-quality raw farmland images are acquired. Size normalization is performed to adjust the image to 512×384 pixels to unify the input scale of global perception analysis and avoid inconsistent feature extraction accuracy due to differences in the resolution of the original image; (x,y) represents the two-dimensional spatial coordinates of the pixel, where x is the index of the pixel in the horizontal width direction of the image and y is the index of the pixel in the vertical height direction of the image.

[0070] In some embodiments, a bilinear interpolation algorithm is used for size adjustment. The algorithm determines the target pixel value by calculating the weighted sum of the gray values ​​of the four adjacent pixels corresponding to the target pixel in the original farmland image. The specific weights can be determined by the inverse ratio of the distance between the target pixel and its adjacent pixels. Finally, the pixel values ​​of the normalized image are normalized.

[0071] Step S120: Convert the RGB color space to the LAB color space and separate the luminance information from the chrominance information;

[0072] In some embodiments, the color space conversion uses the CIE-Lab color model, and its conversion formula is as follows:

[0073]

[0074] In the formula, Indicates the brightness channel. Indicates red and green lanes. Indicates the yellow-blue channel. , , These represent the stimulus values ​​of the three primary colors, red, green, and blue, respectively, in the CIEXYZ color space;

[0075]

[0076] In the formula, Indicates the red channel. Indicates a green channel. This represents the blue channel; the converted LAB space can effectively separate luminance and chromaticity information, facilitating subsequent physical modeling and analysis.

[0077] Step S130: Use a U-shaped Transformer network with embedded multi-dimensional physical prior knowledge to extract multi-scale features and identify degraded regions and feature distributions. The multi-dimensional physical prior knowledge includes atmospheric scattering prior, vegetation spectrum prior, illumination model prior, and depth distance prior.

[0078] In some embodiments, the encoder of the U-shaped Transformer network structure includes four downsampling stages, each consisting of a multi-head self-attention layer and a feedforward neural network layer; the decoder includes four upsampling stages, which achieve upsampling through transposed convolution and fuse with the feature map of the same level as the encoder through skip connections, preserving local detail features, and finally outputting a multi-scale fused feature map.

[0079] In some embodiments, the physical attention mechanism integrates four key priors: atmospheric scattering priors based on Rayleigh and Mie scattering theories; vegetation spectrum priors based on the strong absorption characteristics of chlorophyll at 680 nm and 430 nm; illumination model priors based on a 30° solar altitude angle and a 120° azimuth angle; and depth distance priors that estimate scene depth using atmospheric perspective principles. Embedding these physical priors into the attention mechanism allows the network to focus on physically consistent regions and channels during feature extraction. Specifically, this includes the following dimensions: tensors in the atmospheric scattering priors. This includes: Rayleigh scattering coefficient, Mie scattering coefficient, atmospheric light intensity, visibility range, scattering angle, relative humidity, particle size distribution, wavelength dependence, turbidity coefficient, and path radiation; based on vegetation reflectance spectral characteristics, the vegetation spectral prior tensor. This includes: chlorophyll red light absorption, chlorophyll blue light absorption, carotenoid absorption, near-infrared reflectance, red edge position, NDVI threshold, leaf area index correlation, water content influence, soil background red light, soil background green light, soil background blue light, and vegetation index; based on optical imaging principles, the tensor in the prior art of the illumination model. Includes: solar altitude angle, solar azimuth angle, sky irradiance, shadow intensity, color temperature, light source spectrum, surface reflectivity, and the Lambertian body assumption; based on perspective projection and atmospheric perspective, the depth distance prior tensor. Includes: normalized focal length, object distance estimation, atmospheric perspective, depth of field, perspective distortion, size-distance correlation; attention weights. middle Representing spatial location using two-dimensional coordinates Flattened out, It represents physical priors, which consist of the four priors mentioned above.

[0080] In this embodiment, the core of the global perception layer is a U-shaped Transformer network integrating a physical degradation model. Upon receiving low-quality farmland images, this layer does not directly modify the images but performs deep analysis based on the physical principle of light attenuation in the air. Through this model, this layer generates three sets of key, physically meaningful prior guidance information, which together constitute a "semantic enhancement blueprint." First, a spatial attenuation heatmap is generated, which quantifies and precisely locates the areas in the image with the most severe blurring and color shift due to atmospheric scattering and absorption. Second, by analyzing the differences in information distribution among the red, green, and blue color channels, a channel priority vector is derived to determine the degree of degradation and the urgency of repair for different color channels. Finally, in a color space more uniform to human vision, the difference between the current image and the ideal target in terms of brightness and color is calculated, thereby generating a set of precise color correction coefficients. This blueprint provides a deterministic physical basis and optimization direction for subsequent dynamic decision-making.

[0081] Step S200: Based on the global perception analysis results, construct a semantic enhancement blueprint that quantifies the physical attributes and optimization requirements of the degraded region;

[0082] Step S200 specifically includes:

[0083] Step S210: Calculate the transmittance distribution based on the physical model of light attenuation, and generate a spatial attenuation feature map of the blurred and color-biased areas;

[0084] Step S210 specifically includes:

[0085] Step S211: Based on atmospheric scattering theory, establish a light attenuation model. The calculation formula for the transmittance diagram can be expressed as:

[0086]

[0087] In the formula, For pixels The transmittance at a point, ranging from 0 to 1, reflects the proportion of light that is not scattered and absorbed by the atmosphere during its propagation from objects in a farmland scene to the image acquisition device. The atmospheric attenuation coefficient for farmland ranges from 0.5 to 1.2 and is determined by the atmospheric composition of the farmland environment. The larger the coefficient, the stronger the attenuation effect of the atmosphere on light. For pixels Scene depth, i.e., pixels The greater the actual distance to the image acquisition device and the greater the depth, the longer the light propagation path and the more significant the attenuation.

[0088] Step S212: Estimate the scene depth distribution using dark channel priors; Estimate using dark channel priors:

[0089]

[0090] in: For pixels The prior of the dark passage; Atmospheric light intensity is a vector containing R, G, and B channel components. Its value selection logic is to select the high-value pixels with the highest pixel values ​​from the dark channel. These pixels correspond to the areas in the original image that are most affected by atmospheric scattering. These high-value pixels are mapped back to the original RGB image, and the average values ​​of their R, G, and B channels are calculated respectively to ensure that the atmospheric light intensity truly reflects the color and brightness characteristics of the ambient atmospheric light. In pixels A rectangular area centered on a local window, used to eliminate single-pixel noise and ensure the smoothness of the dark channel; For the color channels in the original RGB image At pixel Pixel intensity at that location;

[0091] Step S213: Calculate the transmittance distribution by integrating the atmospheric attenuation coefficient of farmland with the scene depth distribution. The spatial attenuation thermal calculation formula is as follows:

[0092]

[0093] In the formula, For pixels Spatial attenuation thermal value, used to quantify the degree of pixel degradation, with a value range of [0,1]. A value >0.7 indicates blurred areas caused by haze in farmland images, and a value >0.5 indicates shadowed areas caused by haze in farmland images. When it approaches 1, A value close to 0 indicates that the pixel has almost no degradation;

[0094] Step S214: Generate a spatial attenuation feature map that identifies highly blurred regions based on the transmittance threshold;

[0095] In some embodiments, spatial attenuation thermograms are obtained by setting a transmittance threshold. Binarization and gradient enhancement: Regions with a value >0.7 are marked as "high-priority degradation regions", and regions with a value <0.5 are marked as "high-priority degradation regions". Regions with a value ≤0.7 are marked as "medium priority degradation regions"; Regions with a value ≤0.5 are marked as "low priority regions," ultimately forming a spatial attenuation feature map that includes the location and priority of the degraded regions, providing a spatial positioning basis for subsequent enhancement.

[0096] Step S220: Determine the degree of channel degradation by analyzing the variance distribution of each color channel, and generate channel priority feature vectors;

[0097] In some embodiments, the variances of the red (R), green (G), and blue (B) channels of the original RGB image are calculated separately. , , The larger the variance, the more drastic the fluctuation in pixel values ​​of that channel, and the more severe the information loss. The degree of channel degradation is determined by analyzing the variance distribution of each color channel, and the channel priority vector is calculated based on variance sensitivity analysis.

[0098]

[0099] in: For channel variance For the first One channel, Number of channels For channel The average pixel value; ;

[0100] Step S230: Calculate the color difference between the current image and the standard farmland in the LAB color space, and generate color correction coefficients;

[0101] Step S230 specifically includes:

[0102] Step S231: Obtain the average luminance and chrominance values ​​of the LAB channels of the current image; specifically, for the LAB image transformed in step S120, traverse all pixel coordinates and calculate the average values ​​of the three channels: the average value of the luminance channel. This reflects the overall brightness level of the image; the average value of the red and green channels. This reflects the red-green bias of the image; the mean of the yellow and blue channels. This reflects the yellow-blue bias in the image.

[0103] Step S232: Calculate the color offset of each channel based on the pre-stored healthy farmland standard value;

[0104] Step S233: Generate color correction coefficients for brightness correction and chromaticity compensation based on the color offset of each channel;

[0105] In some embodiments, the color correction coefficient is calculated using a target-current difference model:

[0106]

[0107] In the formula, This refers to the L-channel correction factor; This is the correction factor for channel A; For channel B, the correction factor is used; when When this occurs, it indicates that the crop area has insufficient green saturation;

[0108] Step S240: Combine the spatial attenuation feature map, channel priority vector and color correction coefficient, and fuse the vegetation mask estimation results to form the semantic enhancement blueprint;

[0109] In some embodiments, semantically enhanced blueprint tensors Defined as:

[0110]

[0111] In the formula, For pixels Spatial priority; For pixels Color correction coefficients; For pixels Vegetation cover;

[0112] In some embodiments, vegetation masking estimation is defined as:

[0113]

[0114] In the formula, Normalized Difference Vegetation Index; This is the NDVI threshold; if it exceeds the threshold, the area is considered vegetation; otherwise, it is considered background.

[0115] In some embodiments, the normalized vegetation index (NVI) is defined as:

[0116]

[0117] In the formula, , These are the green and red channel pixel values ​​of the original RGB image; As a reference constant; , These are the regression coefficients;

[0118] In some embodiments, Space priority is defined as:

[0119]

[0120] In the formula, As the global priority benchmark; Local variance weights; For pixels The variance of the local window reflects the degree of regional degradation. By using this formula, based on the global benchmark, and superimposing correction terms for local variance and weights, severely degraded areas such as the haze center and shadow edges are given higher priority, ensuring the precise allocation of enhanced resources.

[0121] Step S200 achieves precise deconstruction of farmland image degradation features and clarification of enhancement requirements. Based on the spatial attenuation feature map of atmospheric scattering theory, combined with dark channel priors and transmittance calculations, the physical rationality of locating degradation areas such as haze and shadows is ensured, avoiding misjudgment of degradation areas by traditional methods. The channel priority vector accurately identifies the most severely degraded color channels, such as the red channel, through variance sensitivity analysis, providing a clear order for differentiated restoration. The color correction coefficients in the LAB space are based on healthy farmland as the standard. Through the correlation calculation between offset and standard values, targeted adjustments to brightness and chromaticity are achieved, avoiding color distortion. The fusion of vegetation mask and spatial priority ensures that the blueprint focuses on the core vegetation areas of the farmland, ensuring that enhancement resources are tilted towards key areas. The final semantic enhancement blueprint transforms the abstract enhancement target into quantitative physical prior information containing degradation location, restoration order, adjustment magnitude, and key areas, providing clear guidance for subsequent reinforcement learning agents to select the optimal enhancement operation, and solving the problems of unclear enhancement targets and lack of physical constraints in existing methods.

[0122] Step S300: Initialize the reinforcement learning agent according to the semantic enhancement blueprint, and select the image processing operation sequence through the physical constraint reward function;

[0123] Step S300 specifically includes:

[0124] Step S310: Based on the spatial decay feature map, channel priority feature vector, and color correction coefficients of the semantic enhancement blueprint, configure the initial state space of the reinforcement learning agent; transform the semantic enhancement blueprint generated in step S200 into a state vector that the reinforcement learning agent can understand, realizing the mapping between physical features and the decision space, core state It can be defined as:

[0125] In the formula, The transmittance in step S213 represents the proportion of light that is not scattered by the atmosphere at this pixel. The lower the value, the more severe the influence of atmospheric attenuation such as haze and shadows on this area. The spatial priority of step S240 is determined by the urgency of degradation in the region where the pixel is located, which provides a basis for prioritizing the enhancement operation. A higher value indicates that the region should be given priority for enhancement processing. The color correction coefficient in step S233 is a vector containing correction parameters for three channels: brightness, red-green, and yellow-blue, which correspond to the brightness adjustment range, red-green saturation offset correction range, and yellow-blue saturation offset correction range in the LAB color space, respectively. The vegetation mask in step S240 is a binarization parameter, where 1 indicates that the pixel belongs to the vegetation area and 0 indicates that the pixel belongs to the non-vegetation background area; this state represents encoding the physical prior information into a feature vector that the reinforcement learning agent can understand, providing physical guidance for subsequent decisions.

[0126] Step S320: Configure an extended action set including physical guidance operations and image processing operations; the physical guidance operations include physical guidance filtering for high attenuation regions, adaptive dehazing operation based on atmospheric scattering model, channel collaborative compensation operation based on vegetation spectral characteristics, LAB color space contrast stretching operation, multi-scale detail fusion operation, Retinex illumination homogenization operation, and adaptive noise suppression operation.

[0127] In some embodiments, the four common image operations include: histogram equalization, which expands the dynamic range of the image and improves the overall contrast by adjusting the pixel value distribution; gamma correction, which adjusts the pixel response through a non-linear function to optimize image brightness; white balance correction, which eliminates color cast and restores true colors by adjusting the RGB channel gain; and linear brightness adjustment, which directly adjusts the overall brightness of the image through a linear scaling factor to adapt to the target brightness range.

[0128] In some embodiments, the physical guided filtering operation is defined as:

[0129]

[0130] In the formula, To output the image in The pixel value at that location reflects the enhanced local contrast.

[0131] The transmittance enhancement factor; This is a Gaussian kernel function; used to simulate the spatial diffusion effect of atmospheric scattering, its mathematical form is: ,in is the standard deviation of the Gaussian kernel, controlling the smoothing range; W is the normalization factor, a standardization coefficient that ensures the pixel values ​​after filtering are within a reasonable range, preventing pixel value overflow after filtering. Its mathematical form is: This operation is for Localized contrast enhancement in the affected area; For the input image in coordinates The original pixel value at that location;

[0132] The adaptive atmospheric scattering removal operation is defined as follows:

[0133]

[0134] In the formula, To the image after descattering The pixel values ​​at that location are used to restore a clear scene; To optimize atmospheric light intensity, This is the lower limit threshold for transmittance. Atmospheric light calibration coefficient; This is the original transmittance; this operation improves the quality of degraded scene restoration by transmittance truncation and atmospheric light calibration.

[0135] Vegetation corridor collaborative compensation operation is defined as:

[0136]

[0137] In the formula, This is the visible light compensation coefficient. For near-infrared compensation coefficient, As a mulch for vegetation, Element-wise multiplication; , To compensate for the pixel values ​​in the red and green channels and enhance the spectral characteristics of vegetation; To estimate the near-infrared channel pixel values, this operation enhances the red light response and chlorophyll characteristics of the vegetation zone using a spectral mixing model;

[0138] The LAB color space contrast stretching operation is defined as follows:

[0139]

[0140] In the formula, , , These are the lightness, red-green, and yellow-blue channel values ​​for the stretched LAB space; The brightness stretch factor. As a red-green color modulator, This is a blue-yellow tint adjustment factor. This operation achieves decoupled contrast enhancement in a perceptually uniform color space;

[0141] Multi-scale detail fusion operation is defined as:

[0142]

[0143] In the formula, To enhance multi-scale details in the fused output image; For the first of Laplace's pyramids Layer, its mathematical expression is ,in This represents the i-th level of the Gaussian pyramid. The pyramid reconstruction operator is mathematically expressed as follows: The Laplace pyramid is constructed by downsampling layer by layer to create a Gaussian pyramid, and then the differences between adjacent layers are calculated to form each layer of the Laplace pyramid. The reconstruction operator is implemented by upsampling and accumulation layer by layer. B is the guiding feature map. Enhanced weighting for details. This operation improves texture saliency by guiding enhanced frequency domain fusion;

[0144] The Retinex illumination homogenization operation is defined as follows:

[0145]

[0146] In the formula, Gaussian kernel , The reflectivity compression factor ( ), To achieve a uniform illumination field, this operation suppresses shadows through illumination-reflection decoupling.

[0147] The adaptive noise suppression operation is defined as follows:

[0148]

[0149] In the formula, Noise mask ( ), For bilateral filtering, The standard deviation of colorimetry ( This operation maintains noise suppression at the edges in flat areas;

[0150] Step S330: Construct a state-action value function based on the Q-learning algorithm, where the state representation includes the fusion features of the current image features and semantically enhanced blueprint information; specifically, this step constructs a decision model based on the Q-learning algorithm, and guides the reinforcement learning agent to select the optimal image processing operation sequence through a physical constraint reward function. The state is defined as the fusion features of the current image's pixel-level features and semantically enhanced blueprint information; the action is defined as an operation in the extended action set constructed in step S320.

[0151] In some embodiments, the reward function consists of a base reward and a blueprint alignment reward, and can be represented as:

[0152]

[0153] In the formula, the reward function consists of a "basic reward" and a "blueprint alignment reward"; The total reward value is calculated by taking the difference between the image quality evaluation metrics UIQM, UCIQE, and UCD before and after execution. The larger the difference, the more significant the improvement in image quality the action has. This is an adjustment factor used to balance the weights of base rewards and blueprint alignment rewards. The blueprint alignment reward is calculated based on the alignment degree. The higher the alignment degree, the better the match between the action and the semantically enhanced blueprint. For quality increment; the alignment calculation formula is:

[0154]

[0155] In the formula, The action performed at time t; For semantic enhancement blueprints, for Norm;

[0156] In some embodiments, the reinforcement learning agent, based on the current state S, traverses all actions A in the extended action set and calculates the action value function value for each action; selects the action with the largest action value function value to execute, and records the physical basis of the action; after executing the action, calculates the base reward and alignment reward to obtain the total reward R; updates the action value function value according to the update rule of Q learning; repeats the above steps until the preset termination condition is met, and finally forms a complete optimal image processing operation sequence.

[0157] In some embodiments, the processing priority order of each color channel is determined according to the channel priority feature vector; step S220 obtains the channel priority feature vector, for example, it is found that the red light channel is affected by haze and other factors, the pixel value is highly discrete and the information is seriously lost, and it is determined that its degradation is the most serious, so the processing priority of the red light channel is set to the highest in subsequent image processing operations.

[0158] In some embodiments, processing intensity parameters for different regions are set based on the spatial distribution characteristics of the spatial attenuation feature map. The spatial attenuation feature map generated in step S210 identifies the degree of degradation in different regions of the image caused by atmospheric scattering, insufficient illumination, etc. The feature map can be divided into regions, and processing intensity parameter rules can be preset for different region types. For example, for highly degraded regions, the processing intensity is increased when performing operations such as descattering and contrast enhancement to powerfully repair severe degradation.

[0159] In some embodiments, the correction intensity parameter for each color channel is set based on the color correction coefficient; based on the color correction coefficient in step S230, the color correction coefficient is analyzed. A large absolute value of the correction coefficient indicates that the channel has a severe color cast. In this case, when performing the color correction operation, a high correction intensity parameter is set for that channel to increase the correction intensity.

[0160] In some embodiments, the image processing operation sequence is selected by a physical constraint reward function, including: calculating the spatial score of the operation effect and the spatial decay feature map, calculating the increment of the image quality evaluation index relative to the previous operation; after each image processing operation, selecting a suitable quality evaluation index for the farmland image; calculating the difference between the image index value after the current operation and the value before the operation, i.e., the quality increment, which reflects the improvement effect of the operation on image quality. The larger the increment, the more effective the operation is in visual effect optimization.

[0161] In some embodiments, calculating the matching degree between the current operation and the semantic enhancement blueprint includes: calculating the consistency between the operation effect and the spatial distribution of the spatial decay feature map; the spatial decay feature map in the semantic enhancement blueprint clarifies the spatial distribution of the degradation region. The image after the current operation is applied is compared with the original spatial decay feature map to analyze whether the repair of high, medium, and low degradation regions by the operation matches the feature map indication; calculating the consistency between the operation effect and the channel processing order of the channel priority feature vector; checking whether the current operation follows the processing order determined by the channel priority feature vector; calculating the consistency between the operation effect and the color correction effect of the color correction coefficients, comparing the actual correction results of each color channel after the current operation with the expected correction direction and magnitude of the color correction coefficients in step S230;

[0162] In summary, the dynamic decision-making layer provided in step S300 serves as the core execution unit, employing a reinforcement learning-based intelligent agent to automatically construct the optimal sequence of image enhancement operations. Its action set not only includes various traditional image processing operations such as contrast adjustment and sharpening, but also innovatively introduces three physically guided operations led by the global perception layer, such as targeted filtering based on attenuation heatmaps. The agent's decision-making process is guided by a composite reward function. This function rewards behaviors that significantly improve general image quality evaluation metrics, as well as operations that are highly aligned with the physical blueprint, thus ensuring that each decision is both effective and physically plausible.

[0163] Step S400: Execute the image processing operation sequence and terminate the enhancement process based on the incremental threshold of the quality assessment index and the physical consistency of the semantic enhancement blueprint;

[0164] Step S400 specifically includes:

[0165] Step S410: Calculate the mean increment of the image quality assessment index for multiple operations; continuously monitor the changes in the image quality assessment index during the process of the reinforcement learning agent performing the image enhancement operation sequence;

[0166] In some embodiments, two typical quality assessment metrics are used: the Universal Image Quality Index (UIQM), which comprehensively considers the brightness, contrast, entropy and other characteristics of an image, and the higher the value, the better the image quality; and the Unreferenced Image Quality Assessment Index (UCIQE), which is designed for natural images and focuses on color saturation, contrast and brightness uniformity, and the higher the value, the better the visual effect.

[0167] Step S420: Calculate the physical consistency alignment between the current operation sequence and the semantic enhancement blueprint, the alignment being calculated based on the consistency between the current operation sequence and the spatial decay feature map, the channel priority feature vector, and the color correction coefficient;

[0168] Step S430: Terminate the process when the mean increment is lower than the first threshold and the alignment is higher than the second threshold.

[0169] In some embodiments, the termination condition determination employs a dual threshold constraint:

[0170]

[0171] in: and This represents the average increment of the last three operations; This means that the average increase in UIQM over the last three operations is extremely low, and the image quality is difficult to improve through subsequent operations. That is, the average increase of UCIQE in the last 3 operations is extremely low, and there is very little room for improvement in visual effect; For the current blueprint alignment, The current operation sequence has a very high physical consistency with the semantic enhancement blueprint, and continuing the operation is likely to introduce over-enhancement;

[0172] Step S400 achieves intelligent control of the enhancement process through a dual-threshold termination mechanism. When the quality index increment tends to stabilize and the physical consistency reaches a high standard, the processing is automatically terminated. This avoids image distortion and waste of computing resources caused by over-enhancement, while ensuring a high degree of consistency between the output result and the physical blueprint. Thus, while ensuring the enhancement effect, the processing efficiency is significantly improved, providing a reliable and adaptive termination criterion for farmland image enhancement.

[0173] Step S500: Output an interpretability report of the physical basis for tracing the enhanced image and associated image processing operation sequence;

[0174] In some embodiments, the algorithm for generating physical parameter traceability reports is as follows:

[0175]

[0176] Where Report is the final generated interpretability report; n is the total number of operations performed in step S300 of reinforcement learning; It is the union of all operations from k=1 to n, that is, a summary of all operation records; k is the operation sequence index, from 1 to n; The type of operation for the k-th operation, such as adaptive atmospheric scattering removal or vegetation channel compensation, is derived from the action space of step S300. For the parameters of the k-th operation, such as the atmospheric light calibration coefficient for descattering operation and the lightness stretching factor for color correction; The interpretation function for the k-th operation associates the operation parameters with physical principles, for example:

[0177] This can be understood as: if the k-th operation is to enhance the contrast in the spatial decay region, then Note: "In the decay heatmap, the region..." The attenuation value is greater than 0.7, so a contrast enhancement of intensity 0.52 is performed on the region in response to the high degradation indication of the attenuation heatmap;

[0178] That is, if the k-th operation is green channel compensation, then Note: In the channel priority vector, the green channel priority is greater than 0.3, which is used to match the channel priority judgment.

[0179] In some embodiments, for each operation performed by the reinforcement learning decision layer in step S300, the self-explanatory output layer synchronously records the operation type and parameters, that is, clearly records which type of operation, such as "transmittance compensation" or "color correction," and the specific parameters; physical basis tracing: associating the physical features of step S200 with explaining why the operation was performed, such as "due to regional..." A decay value > 0.7 indicates a high degradation requirement in the response space decay heatmap;

[0180] Summarize all operation logs from step S300 and integrate them according to the above structure using the formula:

[0181] The first column (Operation): lists the type of each operation;

[0182] The second column (Parameter) lists the parameters for each operation;

[0183] The third column (PhysicalBasis) explains the physical origins of the operation by interpreting the function, relating it to the spatial attenuation heatmap, channel priority, and color correction coefficients.

[0184] Finally, the enhanced image output from step S400 is packaged together to form a complete report containing an "enhancement step map + physical parameter tracing". This method achieves interpretability and transparency throughout the enhancement process by generating a detailed report tracing the physical basis of the operation sequence. It not only outputs high-quality enhanced images but also provides the decision-making basis and physical parameter sources for each processing step, transforming the entire enhancement process from a "black box" to a "white box". This significantly improves the credibility and acceptability of the processing results, providing necessary interpretability guarantees for the practical application of farmland image enhancement technology. In this embodiment, the self-interpretive output layer and the decision layer work in parallel, accurately tracking and recording every action performed by the decision layer to form a detailed operation log. This log includes not only the type and parameters of the operation but also clearly indicates the physical source upon which the operation is based, such as whether it is in response to an attenuation heatmap or based on channel priority. Finally, while outputting the enhanced image, this layer generates a complete interpretability report containing a detailed enhancement step map and physical parameter tracing information, providing a scientific and verifiable physical explanation for each enhancement step.

[0185] Please see Figures 2 to 4 As shown, the following is an enhancement process performed on a cornfield image taken under hazy weather conditions in a plain. The image resolution is 1024×768 pixels, and the visibility at the time of shooting was approximately 2 kilometers. The specific implementation process of the present invention is described in detail below:

[0186] Image preprocessing: Input low-quality farmland images Next, the image was first normalized to 512×384 pixels. Then, the RGB image was converted to the LAB color space, where L∈[0,100] represents luminance information, and a∈[-127,127] and b∈[-127,127] represent chrominance information. The specific conversion process included: normalizing the RGB values ​​to the range [0,1], applying the sRGB to XYZ conversion matrix, and then converting to the CIE-Lab space. This conversion effectively separated luminance and color, laying the foundation for subsequent physical modeling and analysis.

[0187] Global Perception Analysis: A U-shaped Transformer network is used to extract multi-scale features from the input image. The network encoder consists of four downsampling stages, each integrating a multi-head self-attention mechanism and a feedforward network; the decoder corresponds to four upsampling stages, fusing multi-scale features through skip connections. The physical attention mechanism integrates four key priors: atmospheric scattering priors based on Rayleigh and Mie scattering theories; vegetation spectrum priors based on the strong absorption characteristics of chlorophyll at 680nm and 430nm; illumination model priors based on a 30° solar altitude angle and a 120° azimuth angle; and depth distance priors based on atmospheric perspective principles to estimate scene depth.

[0188] Transmittance map calculation: Transmittance map calculation based on atmospheric scattering model: Under typical haze conditions, the atmospheric attenuation coefficient β for farmland is taken as 0.8, and the scene depth is... Estimated through dark channel priors: The calculation results show that the transmittance is approximately 0.3-0.5 in the distant area (distance > 100m), approximately 0.5-0.7 in the mid-range area (distance 50-100m), and approximately 0.7-0.9 in the near-field area (distance < 50m). Based on this, a spatial attenuation heat map was generated. (Value range [0,1]), where areas with a value >0.7 are identified as areas requiring intensive defogging treatment.

[0189] Channel priority calculation: Optimization focus is determined by analyzing the degradation level of each color channel: Red channel variance. Green channel variance Blue channel variance Based on this, the channel priority vector is calculated. The results indicate that the green light channel has the richest information, the blue light channel has the highest priority, and the blue light channel is the most severely degraded. This result is consistent with the physical law that blue light scattering is strong under hazy conditions.

[0190] Color correction factor calculation: Calculate the difference between the current image and the healthy farmland standard in the LAB space: Current image Health standards Based on this, the color correction coefficient is derived. This indicates a need to increase brightness by 21.7%, enhance red-green contrast by 80.7%, and enhance yellow-blue contrast by 96.9%.

[0191] Semantic augmented blueprint generation: The above three sets of information are merged to form a semantic augmented blueprint.

[0192]

[0193] The NDVI of vegetation cover was estimated using NDVI, with a typical NDVI value of 0.65 for healthy cornfields.

[0194] Reinforcement learning enhancement processing: Initialize the Q-learning agent with an action set containing 11 operations. The agent selects an action sequence based on the current state and blueprint information: First, it executes the "Adaptive Atmospheric Scattering Removal" operation, setting the atmospheric light intensity A=220 and the lower limit of transmittance. This improved the UIQM index from 1.24 to 1.78; subsequently, a "vegetation corridor synergistic compensation" operation was performed, applying the visible light compensation coefficient. and near-infrared compensation coefficient The red light response of the vegetated area was enhanced, increasing the UCIQE index from 0.52 to 0.71; finally, the "LAB color space contrast stretching" operation was performed, using a lightness stretching factor. Color adjustment factor , This improved the UCD index from 0.38 to 0.59.

[0195] Termination Criterion: After 8 iterations, the quality indicator increment ΔUIQM = 0.02 is lower than the threshold of 0.05, indicating blueprint alignment. If the threshold of 90% is met, the termination condition is met, and the enhancement process ends.

[0196] Output Enhancement Results: The final output includes an enhanced image of the cornfield and an interpretability report. Enhancement effect evaluation shows that UIQM improved from 1.24 to 2.13 (an improvement of 71.8%), UCIQE from 0.52 to 0.84 (an improvement of 61.5%), and UCD from 0.38 to 0.67 (an improvement of 76.3%). Detailed physical parameter traceability records the processing basis: Atmospheric scattering removal, based on an attenuation model with β=0.8, removed 67% of the atmospheric veil; vegetation compensation enhanced the red light channel by 30% based on chlorophyll absorption characteristics; color correction, based on LAB spatial difference analysis, improved brightness by 20% and enhanced color saturation by 80%.

[0197] In summary, the present invention has at least the following significant advantages over the prior art:

[0198] This invention represents the first successful transformation of farmland image enhancement from a "black box" to a "white box," with each processing step possessing a clear physical basis and traceability. By constructing a meta-enhancement framework based on a physical degradation model, this invention achieves a complete enhancement process from physical perception and dynamic decision-making to interpretable output. This framework transforms traditional "black box" deep learning methods into a "white box" system with a clear physical basis. Through physical model guidance and an operational pipeline tracking mechanism, it provides a scientific and verifiable physical explanation for each enhancement step, solving the key problem of lack of traceability in agricultural applications of deep learning methods.

[0199] This invention employs intelligent adaptive decision-making based on multidimensional physical prior constraints. It can intelligently adjust processing strategies according to different farmland environmental characteristics, exhibiting strong environmental generalization ability and decision robustness. The invention innovatively embeds physical prior knowledge into the state-action space of reinforcement learning, constructing a multidimensional physical constraint mechanism that includes atmospheric scattering priors, vegetation spectrum priors, illumination model priors, and depth-distance priors. This mechanism enables the intelligent agent to adaptively adjust processing strategies and parameters according to different farmland environmental characteristics (such as illumination conditions, atmospheric transparency, and vegetation type), demonstrating stronger environmental generalization ability and decision robustness compared to traditional fixed-process methods.

[0200] This invention employs a physically constrained, high-efficiency convergence optimization approach. By leveraging physical blueprint constraints, it significantly improves training efficiency and convergence stability while avoiding ineffective searches and wasted computational resources. The invention effectively narrows the action search space of the reinforcement learning agent through a physical constraint mechanism based on semantic enhancement blueprints, transforming the unconstrained image enhancement problem into a directed, physically guided optimization problem. This design significantly improves the model's training efficiency and convergence stability, demonstrating a marked improvement in training efficiency compared to traditional reinforcement learning methods, while avoiding the waste of computational resources caused by blind trial and error.

[0201] This invention presents a practical self-supervised learning approach for reference-free evaluation, eliminating the reliance on high-quality reference images and making it particularly suitable for the practical constraints of obtaining standard references in real-world farmland applications. The invention constructs a composite reward mechanism that integrates image quality metrics (UIQM, UCIQE, UCD) and physical consistency constraints, thus eliminating the need for high-quality reference images. This design is particularly well-suited to the practical constraints of obtaining standard reference images in real-world farmland applications, achieving unsupervised image quality evaluation and optimization through self-supervised learning based on a physical model.

[0202] This invention achieves multi-level collaborative processing from pixel-level detail enhancement to global visual effect optimization through cross-scale, multi-space collaborative feature preservation enhancement, effectively balancing local detail preservation with global visual consistency. It utilizes cross-color space operations and a vegetation channel collaborative compensation mechanism to achieve this multi-level collaborative processing. The method performs decoupled contrast enhancement in the LAB perceptually uniform color space, combined with multi-scale Laplacian pyramid fusion technology, effectively solving the technical challenge of existing methods in balancing local detail preservation with global visual consistency when processing complex farmland scenes.

[0203] Please see Figure 5 The diagram illustrates a structural schematic of an interpretable farmland image enhancement system based on physical perception and reinforcement learning, according to an embodiment of the present invention. The system includes:

[0204] The image perception module is configured to acquire raw farmland images and perform global perception analysis, extract multi-scale features, and identify degraded areas and feature distributions.

[0205] The blueprint building module is configured to build a semantically enhanced blueprint that quantifies the physical attributes and optimization requirements of the degraded region based on the global perception analysis results.

[0206] The decision agent module is configured to initialize a reinforcement learning agent based on the semantic enhancement blueprint and select an image processing operation sequence through a physical constraint reward function;

[0207] The enhancement execution module is configured to execute the image processing operation sequence and terminate the enhancement processing based on the incremental threshold of the quality assessment index and the physical consistency of the semantic enhancement blueprint.

[0208] The output module is configured to output an interpretability report that traces the physical basis of the enhanced image and associated image processing operation sequence.

[0209] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0210] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An interpretable farmland image enhancement method based on physical perception and reinforcement learning, characterized in that, The method comprises: Obtaining an original farmland image and performing global perception analysis, extracting multi-scale features and identifying degraded areas and feature distribution; Based on the global perception analysis result, a semantic enhancement blueprint for quantifying the physical properties of the degraded area and optimizing the demand is constructed, including: Based on the light attenuation physical model, the transmittance distribution is calculated to generate a spatial attenuation feature map identifying the blur and color cast areas; Determine the channel degradation degree by analyzing the variance distribution of each color channel to generate a channel priority feature vector; Calculate the color difference between the current image and the standard farmland in the LAB color space to generate a color correction coefficient; Combine the spatial attenuation feature map, channel priority vector and color correction coefficient, and fuse the vegetation mask estimation result to form the semantic enhancement blueprint; According to the semantic enhancement blueprint, initialize the reinforcement learning agent, and select the image processing operation sequence through the physical constraint reward function; Execute the image processing operation sequence, and terminate the enhancement processing according to the quality evaluation index increment threshold and the physical consistency of the semantic enhancement blueprint; Output the enhanced image and the explainability report of the associated image processing operation sequence physical basis traceability.

2. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 1, wherein, Obtaining an original farmland image and performing global perception analysis, extracting multi-scale features and identifying degraded areas and feature distribution, comprising: Performing size normalization processing on the original farmland image; Convert the RGB color space to the LAB color space to separate the brightness information and the chroma information; Use the U-shaped Transformer network embedded with multi-dimensional physical prior knowledge to extract multi-scale features and identify degraded areas and feature distribution, including atmospheric scattering prior, vegetation spectrum prior, light model prior and depth distance prior.

3. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 1, wherein, Based on the light attenuation physical model, the transmittance distribution is calculated to generate a spatial attenuation feature map identifying the blur and color cast areas, including: Establish a light attenuation model based on atmospheric scattering theory; Estimate the scene depth distribution through the dark channel prior; Calculate the transmittance distribution by fusing the farmland atmospheric attenuation coefficient and the scene depth distribution; Generate a spatial attenuation feature map identifying high blur areas according to the transmittance threshold.

4. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 1, wherein, In the LAB color space, calculate the color difference between the current image and the standard farmland to generate a color correction coefficient, including: Obtain the brightness and chroma average of the current image LAB channel; Calculate the color offset of each channel based on the pre-stored standard value of healthy farmland; Generate a color correction coefficient for brightness correction and chroma compensation according to the color offset of each channel.

5. The physically perceptual and reinforcement learning based interpretable agricultural field image enhancement method of any one of claims 1 to 4, wherein, Initializing the reinforcement learning agent according to the semantic enhancement blueprint, comprising: Based on the spatial attenuation feature map, channel priority feature vector and color correction coefficient of the semantic enhancement blueprint, configure the initial state space of the reinforcement learning agent; Configure an extended action set containing physical guidance operations and image processing operations; Based on the Q learning algorithm, construct a state-action value function, wherein the state represents the fusion features containing the current image features and the semantic enhancement blueprint information.

6. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 5, wherein, Initializing the reinforcement learning agent according to the semantic enhancement blueprint, further comprising: Determine the processing priority order of each color channel according to the channel priority feature vector; The processing intensity parameters of different regions are set based on the spatial distribution characteristics of the spatial decay feature map; The correction intensity parameters of each color channel are set based on the color correction coefficient.

7. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 5, wherein, The image processing operation sequence is selected through a physical constraint reward function, including: The spatial distribution consistency of the operation effect and the spatial decay feature map is calculated. The matching degree of the current operation and the semantic enhancement blueprint is calculated, including: The spatial distribution consistency of the operation effect and the spatial decay feature map is calculated. The channel processing order consistency of the operation effect and the channel priority feature vector is calculated. The color correction effect consistency of the operation effect and the color correction coefficient is calculated. The quality increment and the matching degree are combined to generate a reward value through weighted fusion, which drives the reinforcement learning agent to select the optimal action.

8. The physical perception and reinforcement learning based interpretable agricultural field image enhancement method of claim 5, wherein, The enhancement processing is terminated according to the quality evaluation index increment threshold and the physical consistency of the semantic enhancement blueprint, including: The average increment of the image quality evaluation index of multiple operations is calculated. The physical consistency alignment degree of the current operation sequence and the semantic enhancement blueprint is calculated, which is calculated based on the consistency of the current operation sequence and the spatial decay feature map, the channel priority feature vector, and the color correction coefficient. The processing is terminated when the average increment is lower than a first threshold and the alignment degree is higher than a second threshold.

9. An interpretable agricultural field image enhancement system based on physical perception and reinforcement learning, characterized in that, The system includes: An image perception module configured to acquire an original farmland image and perform global perception analysis, extract multi-scale features, and identify degraded regions and feature distributions; A blueprint construction module configured to construct a semantic enhancement blueprint quantifying the physical properties of degraded regions and optimization requirements based on the global perception analysis results; A decision agent module configured to initialize a reinforcement learning agent according to the semantic enhancement blueprint and select an image processing operation sequence through a physical constraint reward function; An enhancement execution module configured to execute the image processing operation sequence and terminate the enhancement processing according to the quality evaluation index increment threshold and the physical consistency of the semantic enhancement blueprint; An explanation output module configured to output an enhanced image and an explainable report tracing the physical basis of the associated image processing operation sequence.

Images