Intelligent regulation and control system for morel bionic cultivation based on image recognition

Abstract

The invention discloses an intelligent regulation and control system for morel bionic cultivation based on image recognition, particularly relates to the technical field of agricultural intelligent cultivation, and aims to solve the problem of unstable recognition caused by water film highlight, local shielding and time sequence drift. The method comprises the following steps of: inhibiting reflection and shadow interference through an imaging and sorting unit to generate a stable image frame, constructing a bed grid and positioning candidate seeds through a grid seed taking unit, segmenting and tracking a pileus trajectory through an instance tracking unit, calculating morphological evolution and wet mark characteristics through a stage calibration unit to generate a grid-level indicating quantity, and performing classification and classification on the pileus trajectory. The consistency write-back unit verifies consistency after regulation and control and calibrates organic cooperation of parameters, and a closed-loop regulation and control link driven by visual evidence is achieved; combined constraints are established among inhibition enhancement, instance segmentation and cross-frame re-identification, and steady-state calibration is introduced into grid mapping, so that monomer positions, forms and growth stage extraction are more accurate, and regulation and control actions are matched with actual requirements of a bed surface, thereby improving stability and efficiency of morel bionic cultivation.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural cultivation technology, and more specifically, to an intelligent control system for the simulated ecological cultivation of morel mushrooms based on image recognition. Background Technology

[0002] Ecological morel cultivation is mostly carried out in shaded sheds or under forests, with the bed surface consisting of soil, humus, and dead leaves and twigs, creating a humid environment with variable light. Spraying and nighttime condensation often form a water film on the caps and mud surface, resulting in alternating specular reflections and backlighting. The arched frame and shading fabric create striped shadows and localized shading. Inspection cameras collect continuous images from multiple perspectives at close range along the bed surface, aiming to extract the location, morphology, and growth stage of individual mushrooms through image recognition, mapping these stage indicators to a bed grid to drive humidification, ventilation, and supplemental lighting, forming a closed-loop control chain based on visual evidence. Current recognition paths mostly originate from controlled environments, often relying on threshold segmentation, single-frame detection, and color proportion. While these processes run smoothly under regular targets and uniform backgrounds, in ground-planted scenarios, they need to consider water film reflections, similar-colored mud surfaces, and complex textures, and also withstand the disturbances to temporal stability caused by changes in perspective and scale.

[0003] However, the core contradictions are concentrated in three areas. First, the high contrast of the water film and the similar color of the mud surface and the cap make single-frame segmentation prone to boundary adhesion and breakage, making it difficult to reliably distinguish between primordia and young mushrooms, and lacking reliable basis for stage identification. Second, fallen leaves and skeletal shadows cause local occlusion, and changes in viewpoint and distance bring about scale and perspective drift, making the identity of instances easily drift in continuous images, resulting in unstable re-identification, and stage labels repeatedly jumping within a short period of time. Third, the mapping between visual indicators and bed grids is sensitive to fluctuations, and the recognition noise, after being amplified by thresholds and rules, causes trigger point jitter, leading to mismatch in the timing of local humidification and ventilation, and the closed-loop control exhibits discontinuity and deviation. The problem is not a single threshold inaccuracy, but rather an instability in the link caused by the intertwining of imaging physics, background texture, and temporal correlation. It is urgent to establish joint constraints between suppression enhancement, instance segmentation, and cross-frame re-identification, and to introduce steady-state calibration in the grid mapping stage so that the image recognition results can continuously support the intelligent control of biomimetic cultivation.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent control system for morel mushroom biomimetic cultivation based on image recognition. This system achieves a closed-loop control chain driven by visual evidence through a combination of: an imaging centering unit to suppress reflection and shadow interference and generate stable image frames; a grid seed collection unit to construct a bed grid and locate candidate seeds; an instance tracking unit to segment and track cap trajectories; a stage calibration unit to calculate morphological evolution and wett mark features to generate grid-level indicators; and a consistent write-back unit to verify post-control consistency and calibrate parameters. Furthermore, it establishes joint constraints between reflection suppression enhancement, instance segmentation, and cross-frame re-identification, and introduces steady-state calibration in grid mapping, making the extraction of individual plant positions, morphology, and growth stages more accurate. The control actions match the actual needs of the bed surface, thereby improving the stability and efficiency of morel mushroom biomimetic cultivation and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An intelligent control system for the simulated ecological cultivation of morel mushrooms based on image recognition includes: Imaging and normalization unit: Acquires multi-view sequences of the biomimetic bed surface and generates reflection normalized frames; Mesh seeding unit: Align the reflection normalized frame geometry and construct the bed mesh. Generate candidate seeds in each mesh only in the intersection area of ​​the pore orientation field and the fine-grained texture density map, and bind the mesh coordinates. Instance tracking unit: Starting from the candidate seed, it performs instance segmentation to obtain the instance segmentation boundary and establishes the instance trajectory. It maintains the continuity of identity under the grid coordinate constraint according to the identity preservation rule and fills in the missing segments when there is occlusion and viewpoint change. Stage calibration unit: Calculates morphological evolution along the instance trajectory and generates stage labels based on stage label rules in conjunction with neighboring wet marks. The stage labels are aggregated into grid-level indicators and output trigger guidelines. Consistent write-back unit: In the new image acquired after triggering, the short-term increment of the porosity direction and the shrinkage of the wet mark are extracted as consistency signals. If the consistency signal is inconsistent with the stage label, the candidate seed generation rule, instance segmentation boundary parameters, identity preservation rule, and stage label rule are written back synchronously, and the grid coordinate reference is used during the write-back.

[0007] In a preferred embodiment, the imaging normalization unit acquires multi-view image sequences of the biomimetic bed surface, generates an original image sequence, and sequentially performs reflection suppression on each frame to reduce the intensity of the water film highlight area while preserving texture details, generating a suppressed reflection image sequence; performs shadow correction on the suppressed reflection image sequence, adjusting local brightness to eliminate the influence of striped shadows, generating a shadow-corrected image sequence; performs geometric alignment on the shadow-corrected image sequence, and corrects scale and perspective drift based on a reference frame by feature point matching, generating an aligned image sequence; calculates the temporal median of the aligned image sequence according to pixel position and channel, generating a reflection normalized frame with uniform brightness and clear texture.

[0008] In a preferred embodiment, the grid seed-taking unit receives a reflection normalized frame, aligns it to the bed surface reference coordinate system to generate an aligned reflection frame, constructs a bed grid to divide the bed surface into regular grid units, extracts the pore orientation field to obtain the cap texture direction features, generates a fine-grained texture density map to highlight the density of cap details, generates candidate seeds based on the intersection area of ​​the pore orientation field and the fine-grained texture density map within each grid unit and binds them to the grid coordinates, and outputs a set of candidate seeds for use by the instance tracking unit.

[0009] In a preferred embodiment, the stage calibration unit receives a set of instance trajectories and an original image sequence, calculates the morphological evolution features of the instance trajectories to extract the cap growth dynamics, extracts the neighborhood wet mark features of the bed grid units to reflect the humidity distribution, combines the morphological evolution features and the neighborhood wet mark features to generate stage labels to identify the cap growth stage, aggregates the stage labels as grid-level indicators, and outputs a set of grid-level indicators for use by the consistent write-back unit.

[0010] In a preferred embodiment, the aperture orientation field is obtained based on local texture analysis of aligned reflection frames: for each pixel position, a fixed-size window is selected with the pixel as the center, and the variation of the intensity of the red, green and blue channels in the horizontal and vertical directions within the window is calculated to determine the intensity gradient direction angle of each pixel; by statistically analyzing the distribution of all direction angles within the window, the direction angle with the highest frequency of occurrence is selected as the aperture orientation field value of the corresponding pixel.

[0011] In a preferred embodiment, the acquisition of neighborhood wet stain features is based on pixel analysis of the original image sequence within the bed grid cell: for each frame of image, wet stain regions are identified within the pixel region of each bed grid cell. A wet stain region is defined as a set of pixels with pixel intensity below a wet stain intensity threshold and blue channel intensity higher than red and green channel intensity. The wet stain intensity threshold is determined through sample statistics. The proportion of the number of pixels in the wet stain region to the total number of pixel regions in the grid cell is calculated as the neighborhood wet stain feature.

[0012] In a preferred embodiment, the consistent write-back unit receives a new image sequence and a set of grid-level indicators, extracts the short-term increment of the porosity orientation and wett shrinkage features of the bed grid cells, verifies the consistency with the grid-level stage labels, calibrates the candidate seed generation rules of the grid seeding unit, the instance segmentation boundary parameters and identity preservation rules of the instance tracking unit, and the stage label rules of the stage calibration unit for inconsistent bed grid cells, outputs a set of calibration parameters and uses the center coordinates of the bed grid cells as references.

[0013] The technical effects and advantages of the intelligent control system for morel mushroom biomimetic cultivation based on image recognition in this invention are as follows: This invention achieves a closed-loop control chain driven by visual evidence through the organic coordination of an imaging normalization unit to suppress reflection and shadow interference and generate stable image frames, a grid seed collection unit to construct bed grids and locate candidate seeds, an instance tracking unit to segment and track cap trajectories, a stage calibration unit to calculate morphological evolution and wet mark features to generate grid-level indicators, and a consistency write-back unit to verify consistency after regulation and calibrate parameters. By establishing joint constraints between reflection suppression enhancement, instance segmentation, and cross-frame re-identification, and introducing steady-state calibration in grid mapping, it solves the recognition instability problems caused by water film highlights, local occlusion, and temporal drift, making the extraction of individual location, morphology, and growth stage more accurate. Regulation actions such as humidification, ventilation, and supplemental lighting are time-matched to the actual needs of the bed surface, thereby significantly improving the stability and efficiency of morel mushroom biomimetic cultivation. It avoids the boundary adhesion, identity jumps, and trigger jitter of traditional methods in uncontrolled environments. The synergistic effect of the overall technical logic ensures increased cultivation yield and optimized resource utilization, and provides a scalable intelligent framework for similar complex agricultural scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the intelligent control system for the simulated ecological cultivation of morel mushrooms based on image recognition, as described in this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Figure 1 The present invention provides an intelligent control system for the simulated ecological cultivation of morel mushrooms based on image recognition, comprising: Imaging and normalization unit: Acquires multi-view sequences of the biomimetic bed surface and generates reflection normalized frames, performs reflection suppression and shadow correction for subsequent geometric alignment and texture analysis.

[0017] Mesh seeding unit: Align the reflection normalized frame geometry and construct the bed mesh. Generate candidate seeds in each mesh only in the intersection area of ​​the pore orientation field and the fine-grained texture density map, and bind the mesh coordinates.

[0018] Instance tracking unit: Starting from the candidate seed, it performs instance segmentation to obtain the instance segmentation boundary and establishes the instance trajectory. It maintains the continuity of identity under the grid coordinate constraint according to the identity preservation rule, and fills in the missing segments when there is occlusion or viewpoint change.

[0019] Stage calibration unit: Calculates morphological evolution along the instance trajectory and generates stage labels based on stage label rules in conjunction with neighboring wet marks. The stage labels are aggregated into grid-level indicators and output trigger guidelines.

[0020] Consistent write-back unit: In the new image acquired after triggering, the short-term increment of the porosity direction and the shrinkage of the wet mark are extracted as consistency signals. If the consistency signal is inconsistent with the stage label, the candidate seed generation rule, instance segmentation boundary parameters, identity preservation rule, and stage label rule are written back synchronously, and the grid coordinate reference is used during the write-back.

[0021] An intelligent control system for morel mushroom biomimetic cultivation based on image recognition achieves precise growth monitoring and environmental control through visual evidence, addressing the challenges of complex environments in morel mushroom biomimetic cultivation. Morel mushrooms, as a high-value edible fungus, are typically cultivated in biomimetic environments under shade structures or in forests. The bed surface consists of soil, humus, and fallen leaves and branches. Environmental characteristics include humid air, variable light, water films formed by misting and nighttime condensation, and striped shadows created by the arched frame and shading fabric. These factors lead to a complex imaging environment, manifested as high contrast compression from the water film, similar colors between the cap and the mud surface, local occlusion, and changing viewing angles, interfering with individual mushroom identification, growth stage determination, and the stability of control actions. Traditional image processing methods rely on threshold segmentation and single-frame detection, which are difficult to adapt to uncontrolled scenarios and are prone to boundary adhesion, identity drift, and trigger jitter. This invention processes multi-view image sequences to construct a closed-loop control link, mapping visual indicators to a bed grid to drive humidification, ventilation, and supplemental lighting actions, thereby improving cultivation efficiency and stability.

[0022] The imaging normalization unit, as the first step in the processing chain, is responsible for processing multi-view sequence images of the biomimetic bed surface and generating reflection-normalized frames suitable for subsequent geometric alignment and texture analysis. Its tasks are to suppress water film highlights and shadows, correct imaging inconsistencies, and provide stable input for subsequent mesh seeding and instance tracking.

[0023] The following details the specific processing logic of the imaging alignment unit: 1.1: Multi-view image sequence acquisition.

[0024] The imaging and normalization unit first receives a sequence of multi-view images, called the raw image sequence, acquired at close range along the simulated ecological bed surface by the inspection camera. Each frame contains pixel coordinates and intensity values ​​for the red, green, and blue channels, with the intensity values ​​normalized to the range of zero to one. The inspection camera moves along the bed surface at fixed time or spatial intervals, acquiring images covering the entire bed area and recording information such as the soil cover, caps, water film, and shadows. Due to camera movement, changes in the viewing angle cause scale and perspective drift, which will be corrected in subsequent steps. This step generates the raw image sequence, providing input data containing complete bed surface information for subsequent processing.

[0025] 1.2: Reflection inhibition.

[0026] In the original image sequence, the water film highlights compress image contrast, interfering with the distinction between the cap and the mud surface. Reflection suppression preserves texture details by reducing the intensity of the highlight areas. For each frame, the brightness variation amplitude of each pixel in the horizontal and vertical directions is calculated, and the square root of the sum of the squares is taken to obtain the local brightness variation field. Highlight areas exhibit high intensity and low brightness variation amplitude. By setting brightness thresholds and brightness variation amplitude thresholds, a highlight mask is generated, with highlight areas marked as one and other areas marked as zero. The pixel intensity of the highlight area is multiplied by the ratio of the average intensity of the non-highlight area to the maximum intensity of that pixel to obtain the suppressed intensity value, while the intensity of the non-highlight area remains unchanged. To avoid division by zero, a small constant is added to the maximum value. The generated suppressed reflection image sequence preserves texture details and reduces the interference of water film highlights on texture analysis.

[0027] The brightness threshold and brightness variation amplitude threshold were determined through statistical analysis of typical image samples from the simulated morel cultivation bed. For example, multiple sets of original image sequences containing water film highlights, caps, and mud surfaces were collected. For each frame, the mean intensity of all pixels in the red, green, and blue channels was calculated, and the high percentile (e.g., 90%) intensity value was selected as the brightness threshold to distinguish between highlight and non-highlight areas. Simultaneously, the local brightness variation field of each frame was calculated. Based on the intensity variation amplitude in the horizontal and vertical directions of pixels, the distribution of the variation amplitude was statistically analyzed, and the low percentile (e.g., 10%) value was selected as the brightness variation amplitude threshold to identify the low gradient characteristics of highlight areas. Statistical analysis of samples ensured that the thresholds adapted to different lighting and bed conditions, enhancing the accuracy of reflection suppression.

[0028] 1.3: Shadow Correction.

[0029] In suppressed-reflection image sequences, the striped shadows created by the skeleton and shading fabric cause uneven local brightness, affecting texture extraction. Shadow correction achieves uniformity by adjusting local brightness. For each frame of the suppressed-reflection image, the average intensity of all pixels in the red, green, and blue channels is calculated as the global brightness mean. The local brightness mean is calculated for the region centered on each pixel using a sliding window of fixed size. Shadow regions typically have lower local brightness means. The correction factor is defined as the ratio of the global brightness mean to the local brightness mean; a small constant is added to the local brightness mean to avoid division by zero. The pixel intensity of each frame is multiplied by the corresponding correction factor to obtain the corrected image. Ensuring brightness uniformity provides a consistent image basis for geometric alignment.

[0030] 1.4: Geometric alignment preparation.

[0031] Due to camera movement, shadow-corrected image sequences exhibit scale and perspective drift, necessitating alignment to a unified reference frame. The first frame in the sequence is selected as the reference frame, and the transformation relationship of each frame relative to the reference frame is calculated. A corner detection method is used to extract salient feature points from each frame, denoted as the corner set. Nearest neighbor matching is used to determine corresponding corner pairs between the reference frame and each frame. Based on these corresponding point pairs, the affine transformation matrix, including translation, rotation, and scaling parameters, is solved by minimizing the sum of squared coordinate errors after transformation. This matrix is ​​applied to each shadow-corrected image frame, and non-integer pixels are filled using bilinear interpolation to generate an aligned image. Ensuring a consistent coordinate system across all frames provides a stable spatial reference for mesh seeding.

[0032] 1.5: Reflection-normalized frame generation.

[0033] The aligned image sequence is integrated into a single image, called a reflection-normalized frame, to support texture analysis. For each aligned image frame in the sequence, the median intensity value over time is calculated based on pixel position and channel, serving as the final intensity value for that pixel. Median calculation filters out temporal noise such as instantaneous illumination changes, preserving the stable texture features of the bed surface and cap. The reflection-normalized frame, with uniform brightness and clear texture, serves as input for subsequent mesh seeding units.

[0034] The imaging normalization unit, through multi-view sequential image acquisition, reflection suppression, shadow correction, geometric alignment preparation, and reflection normalization frame generation, systematically handles interference from water film highlights, striped shadows, and viewpoint drift in the simulated morel cultivation scenario. The original image sequence is processed to generate reflection normalization frames with uniform brightness, clear texture, and aligned coordinates, directly supporting subsequent mesh seed extraction units in extracting the pore orientation field and fine-grained texture density map.

[0035] The imaging normalization unit acquires multi-view image sequences, performs reflection suppression, shadow correction, and geometric alignment to generate reflection-normalized frames with uniform brightness and clear texture, providing stable input for subsequent processing. However, the biomimetic bed surface still has complex textures and local occlusions, requiring the extraction of reliable features from the reflection-normalized frames to construct a bed grid and locate candidate seeds to support instance segmentation and tracking. The grid seeding unit, as the second step in the processing chain, is responsible for geometrically aligning the reflection-normalized frames, constructing the bed grid, and generating candidate seeds based on the porosity field and fine-grained texture density map, providing accurate initial localization for the instance tracking unit.

[0036] The following details the specific processing logic of the grid seed collection unit: 2.1: Geometric alignment of reflection-normalized frames.

[0037] The system receives a reflection-normalized frame generated by the imaging normalization unit. This frame contains pixel coordinates and intensity values ​​for the red, green, and blue channels, normalized to the range of zero to one. To ensure spatial consistency on the cultivation bed, the reflection-normalized frame is aligned to a predefined bed reference coordinate system. This system, based on the physical dimensions of the cultivation bed, is defined as a two-dimensional grid containing regularly distributed grid point coordinates. The system detects the bed boundary corner points (e.g., the four corners of the bed) in the reflection-normalized frame and matches them with corresponding points in the bed reference coordinate system. By minimizing the sum of squared errors after corner coordinate transformation, an affine transformation matrix is ​​calculated, including translation, rotation, and scaling parameters. This affine transformation matrix is ​​applied to the reflection-normalized frame, and bilinear interpolation is used to fill non-integer pixels, generating an aligned reflection frame. The pixel coordinates of the aligned reflection frame are consistent with the physical location of the bed, providing a unified spatial reference for bed grid construction.

[0038] 2.2: Bed grid construction.

[0039] Based on aligned reflection frames, a bed grid is constructed to divide the bed surface into regular regions, facilitating the location of candidate cap sites. The bed grid is a uniformly divided two-dimensional grid, with each grid cell identified by its center coordinates. The size of each grid cell is determined based on the physical dimensions of the cultivation bed and the spacing between strains, ensuring that each grid cell corresponds to a potential strain location. Each grid cell covers a specific pixel region in the aligned reflection frame, with the region's extent defined by the grid cell's center coordinates and its horizontal and vertical half-widths. The bed grid binds each grid cell to a corresponding pixel region, providing a clear spatial division for feature extraction.

[0040] 2.3: Extraction of the pore orientation field.

[0041] The surface of the morel cap exhibits a porous texture, and the orientation of these pores reflects the structural characteristics of the cap, helping to distinguish it from a muddy or withered leaf background. For aligned reflection frames, the pore orientation field is calculated, representing the principal orientation angle of the local texture at each pixel. For each pixel, a fixed-size window is taken centered on it, and the intensity variations of the red, green, and blue channels within the window in the horizontal and vertical directions are calculated to determine the orientation angle of the intensity gradient. By statistically analyzing the distribution of orientation angles within the window, the orientation angle with the highest frequency is selected as the pixel value of the pore orientation field. The pore orientation field highlights the texture orientation characteristics of the cap region, providing a structural basis for candidate seed generation.

[0042] 2.4: Generating fine-grained texture density maps.

[0043] Fine-grained texture density reflects the density of details on the cap surface, contrasting with the rough textures of mud and dead leaves. For aligned reflection frames, a fine-grained texture density map is calculated, representing the intensity of texture details at each pixel. For each pixel, a small window is taken centered on it, and the high-frequency component intensities of the red, green, and blue channels within the window are calculated. The high-frequency components are extracted by performing a frequency domain transformation on the pixel intensities within the window, summing the squared amplitudes of the high-frequency components, and normalizing them to the range of zero to one. The fine-grained texture density map highlights the fine-grained features of the cap region, providing supplementary information for candidate seed generation.

[0044] 2.5: Candidate seed generation and grid coordinate binding.

[0045] Based on the pore orientation field and fine-grained texture density map, candidate seeds are generated within each grid cell of the bed grid to represent potential cap locations. Candidate seeds are the intersection regions of the pore orientation field and the fine-grained texture density map, satisfying the conditions of pore orientation consistency and high-density texture. The average direction of the pore orientation field within each grid cell is calculated, and pixels whose difference from the average direction is less than the orientation consistency threshold and whose fine-grained texture density is greater than the fine-grained texture density threshold are selected as candidate seeds. Each candidate seed is bound to the center coordinates of its respective grid cell, generating a candidate seed set. The candidate seed set contains the seed positions and their grid coordinates for all grid cells, providing accurate initial positioning for instance tracking units.

[0046] The orientation consistency threshold and fine-grained texture density threshold were determined through statistical analysis of typical image samples of the simulated morel cultivation bed. For example, multiple sets of reflection-normalized frames containing caps, mud surfaces, and dead leaves were collected. For each frame, the pore orientation field was calculated, and the distribution of pore orientation within each grid cell was statistically analyzed. The concentration of the orientation distribution (e.g., median and interquartile range) was selected as the orientation consistency threshold to ensure that candidate seeds have a consistent texture orientation. Simultaneously, a fine-grained texture density map was calculated, and the density value distribution between the cap and non-cap regions was statistically analyzed. The high percentile (e.g., 80%) of the density value in the cap region was selected as the fine-grained texture density threshold to distinguish the high-density texture of the cap from the background. This method ensures that the threshold adapts to different bed textures and lighting conditions through sample statistics, improving the accuracy of candidate seed generation.

[0047] The mesh seed-collecting unit extracts reliable cap features from a simulated ecological bed surface with complex textures and partial occlusions through geometric alignment of reflection-normalized frames, bed mesh construction, pore orientation field extraction, fine-grained texture density map generation, and candidate seed generation bound to mesh coordinates, generating a candidate seed set. The joint constraints of the pore orientation field and fine-grained texture density map ensure accurate seed localization, while mesh coordinate binding provides a stable spatial reference for subsequent instance tracking. The candidate seed set directly supports the instance tracking unit in instance segmentation and trajectory establishment, ensuring the accuracy of growth stage discrimination.

[0048] The imaging normalization unit generates reflection-normalized frames to eliminate highlight and shadow interference. The grid seed collection unit further constructs a bed grid and generates a candidate seed set, providing accurate initial cap localization. However, local occlusion and viewing angle changes on the biomimetic bed surface cause the cap's identity to drift in continuous images, requiring instance segmentation and cross-frame tracking to maintain identity continuity. As the third step in the processing chain, the instance tracking unit starts with candidate seeds, performs instance segmentation and trajectory establishment, fills in occlusion gaps, and ensures that subsequent calibration units can accurately determine the growth stage.

[0049] The following details the specific processing logic of the instance tracking unit: 3.1: Candidate seed initialization.

[0050] The system receives a candidate seed set generated by the grid seeding unit. This set contains the initial cap position coordinates within each bed grid unit and the coordinates of its associated grid center. Simultaneously, it receives the raw image sequence from the imaging and normalization unit. Each frame contains pixel coordinates and intensity values ​​for the red, green, and blue channels, normalized to the range of zero to one. Candidate seeds are assigned to each frame, and their coordinates are mapped to the corresponding pixel positions, generating an initial instance set. This initial instance set contains the candidate seed coordinates for each bed grid unit in each frame. This step generates the initial instance set, providing a clear starting point for cap localization in subsequent instance segmentation.

[0051] 3.2: Instance segmentation.

[0052] Based on the initial instance set, instance segmentation is performed on each frame of the image to generate instance segmentation boundaries. For each candidate seed, the cap region is extracted using a region growing method centered on its coordinates. Region growing starts from the candidate seed and gradually expands to neighboring pixels. The expansion conditions are that the difference between the intensity of the neighboring pixels and the intensity of the seed pixels is less than an intensity difference threshold, and the difference between the direction of the pore orientation field and the direction of the seed position is less than a direction consistency threshold. These two thresholds are determined through sample statistics. The boundaries of the segmented regions are extracted to generate instance segmentation boundaries, representing the outline of the cap, accurately describing the cap region, and providing boundary features for trajectory establishment.

[0053] The quantization method for the intensity difference threshold is determined by statistically analyzing the distribution of pixel intensity in the image. For example, the global intensity mean and the maximum intensity difference of the image are calculated, and then a fixed proportion of the maximum intensity difference is selected as the threshold. A specific example: In region growing segmentation, the pixel intensity range for grayscale image samples is calculated to be 0 to 255. 10% of this range is selected as the threshold, i.e., 25.5, to ensure that the intensity difference between the seed pixel and its neighboring pixels is less than this value to expand the region.

[0054] The quantification method for the orientation consistency threshold is determined by analyzing the distribution of the texture orientation field. For example, the local orientation angle histogram of the hole texture orientation field is calculated, and the orientation deviation range around the peak of the histogram is selected as the threshold. Specifically, in texture analysis, if the peak deviation of the orientation angle distribution of an image sample is 15°, 20% of this deviation is selected as the threshold, i.e., 3°, to ensure that the orientation angle difference within the candidate region is less than 3° to maintain consistency.

[0055] 3.3: Instance Trajectory Establishment.

[0056] Based on the instance segmentation boundary set, instance trajectories are constructed on the original image sequence to represent the temporal boundary sequence of caps within each bed grid cell. For each image frame and the next frame, the shape similarity and center coordinate distance of the instance segmentation boundaries are compared. Shape similarity is determined by calculating the maximum distance between two boundary point sets, which is the maximum of the minimum point-to-point distances between the two boundary point sets. Center coordinate distance is the Euclidean distance between the center points of the two boundary regions. Boundary pairs with the minimum shape similarity and center coordinate distance are selected and associated as continuous trajectories of the same instance, recording the temporal changes of the caps in the image sequence.

[0057] 3.4: Identity maintenance constraints.

[0058] To ensure the continuity of instance trajectories, constraints are applied based on the center coordinates of the bed grid cells. An identity preservation rule is defined: if the distance between the boundary center coordinates of an instance trajectory and the center coordinates of its corresponding grid cell is less than a distance threshold in both the current and next frames, the instance retains the same identity. For example, the average offset distance of the cap center coordinates in the image sequence is calculated, and the high percentile of the offset distance distribution is selected as the threshold. If the condition is met, the instance trajectory remains continuous between frames; otherwise, it is marked as a new instance or a lost instance. By updating the instance trajectory set, the continuity of the cap identity under grid coordinate constraints is ensured, reducing identity drift caused by changes in viewpoint.

[0059] 3.5: Fill in occlusions and missing parts.

[0060] To address missing instance segmentation boundaries due to occlusion or viewpoint changes, instance trajectories are completed based on information from the previous frame. If an instance segmentation boundary is missing in the current frame, the corresponding instance segmentation boundary in the previous frame is checked. If the boundary exists in the previous frame, the boundary position in the current frame is predicted based on the center coordinates and pore orientation field of the previous frame boundary. The predicted boundary is a pixel region centered at the center coordinates, with a radius smaller than a region radius threshold, and the difference between the pore orientation and the orientation of the center position in the previous frame is less than an orientation consistency threshold. The radius threshold is quantized by analyzing the size distribution of the cap region, for example, calculating the average diameter of the instance segmentation boundary and then selecting a fixed proportion of the diameter as the threshold. The completed boundary is added to the instance trajectory. This step generates the final instance trajectory set, ensuring trajectory continuity and providing stable cap trajectories for the stage calibration unit.

[0061] The instance tracking unit extracts continuous trajectories of the cap from the candidate seed set and the original image sequence through candidate seed initialization, instance segmentation, instance trajectory establishment, identity preservation constraints, and occlusion completion. The instance segmentation boundary set accurately describes the cap contour, while identity preservation constraints and occlusion completion effectively address local occlusion and viewpoint changes, ensuring trajectory continuity and identity stability. The final instance trajectory set supports the stage labeling unit in performing morphological evolution and stage label generation, improving the accuracy of growth stage discrimination.

[0062] The morphology of the cap changes with the growth stage, and the distribution of wet marks affects stage discrimination. It is necessary to extract morphological features from the instance trajectory and combine them with wet mark information to generate reliable stage labels. The stage labeling unit receives the set of instance trajectories and the original image sequence, calculates morphological evolution, generates stage labels by combining them with neighborhood wet marks, and aggregates them into grid-level indicators to drive precise control actions.

[0063] The following details the specific processing logic of the stage calibration unit: 4.1: Instance trajectory reception.

[0064] The system receives a set of instance trajectories generated by the instance tracking unit. This set contains the temporal instance segmentation boundaries of the caps within each bed grid cell, and each instance trajectory records the boundary changes of the caps in the original image sequence. Simultaneously, the system receives the original image sequence, where each frame contains pixel coordinates and intensity values ​​for the red, green, and blue channels, with the intensity values ​​normalized to the range of zero to one. By associating the instance trajectory set with the original image sequence, the system ensures that each instance segmentation boundary corresponds to the correct image frame.

[0065] 4.2: Morphological evolution calculation.

[0066] Based on the set of instance trajectories, the morphological evolution features of each instance trajectory are calculated to reflect the changing trend of the cap over time. For each instance trajectory, the shape parameters of the instance segmentation boundary for each frame are extracted, including the area and aspect ratio of the boundary region. The area is determined by counting the number of pixels within the boundary, and the aspect ratio is determined by calculating the aspect ratio of the minimum bounding rectangle of the boundary. The morphological evolution features consist of the area change rate and the aspect ratio change rate, which are respectively the ratio of the area of ​​the current frame to the area of ​​the previous frame minus one, and the ratio of the aspect ratio of the current frame to the aspect ratio of the previous frame minus one. If the boundary of the previous frame is missing, the area and aspect ratio of the most recent valid frame are used. The set of morphological evolution features is used to describe the growth dynamics of the cap and provide morphological basis for the generation of stage labels.

[0067] 4.3: Extraction of wet marks in the neighborhood.

[0068] Based on the original image sequence, neighborhood wettage features are extracted from the corresponding grid cell of each instance trajectory to reflect the humidity distribution around the cap. For each image frame, wettage regions are identified within the pixel area of ​​the grid cell. A wettage region is defined as a set of pixels with pixel intensity below a wettage intensity threshold and blue channel intensity higher than red and green channel intensity. The wettage intensity threshold is determined by statistically analyzing the distribution of image pixel intensity; for example, the median blue channel intensity in the wettage samples is calculated, and 80% of the median is selected as the threshold to distinguish wettage pixels from background pixels. The proportion of pixels in the wettage region to the total number of pixels in the grid cell is calculated as the neighborhood wettage feature. The neighborhood wettage feature set provides humidity distribution information, providing an environmental basis for stage label generation.

[0069] 4.4: Stage label generation.

[0070] By combining morphological evolution feature sets and neighborhood wettage feature sets, stage labels are generated for each instance trajectory to identify the growth stage of the cap, such as primordium, juvenile, or mature. Stage labeling rules are based on the area change rate and aspect ratio change rate of morphological evolution features, as well as the area proportion of neighborhood wettage features. If the area change rate is less than a first area threshold and the wettage area proportion is greater than the wettage threshold, it is marked as a primordium. The quantification method for the first area threshold is determined by analyzing the distribution of the cap area change rate in the instance trajectory. For example, statistically analyzing the area change rate samples from the primordium to the juvenile stage, the low percentile value of the distribution is selected as the threshold to define the boundary of the early growth stage. The quantification method for the wettage threshold is determined by statistically analyzing the distribution of the wettage area proportion within the grid cell. For example, analyzing humidity samples from different growth stages, the average of the proportion distribution is selected as the threshold to judge the degree of conformity of humidity conditions to the stage label. If the area change rate is between the first and second area thresholds, the aspect ratio change rate is greater than the aspect ratio change rate threshold, and the wettage area proportion is greater than the wettage threshold, the stage label is marked as a primordium. Mushrooms with a wettmark area ratio greater than the wettmark threshold are labeled as juvenile mushrooms. The second area threshold is quantified by analyzing the distribution of cap area change rates in the statistical instance trajectory. For example, the high percentile value of the distribution is selected as the threshold to identify the transition to the late growth stage, calculating the area change rate samples from juvenile to mature mushroom stages. If the area change rate is greater than or equal to the second area threshold, the aspect ratio change rate is less than or equal to the aspect ratio change rate threshold, and the wettmark area ratio is less than the wettmark threshold, the mushroom is labeled as mature. The aspect ratio change rate threshold is quantified by examining the distribution of cap aspect ratio change rates in the instance trajectory. For example, the median of the change rate distribution is selected as the threshold to assess the significance of morphological extension, statistically analyzing morphological evolution samples from the juvenile stage. The stage label set is used to accurately identify the cap growth stage.

[0071] 4.5: Grid-level indicator aggregation.

[0072] Based on the set of stage labels, stage labels are aggregated within each bed grid unit to generate grid-level indicators that drive control actions such as humidification, ventilation, or supplemental lighting. For each bed grid unit, the stage labels of all instance trajectories within it are statistically analyzed, and the label with the highest frequency is selected as the grid-level stage label. The grid-level indicator consists of the grid-level stage label and its confidence level, where the confidence level is the proportion of that label in the instance trajectories within the grid unit. The set of grid-level indicators provides control triggering guidance for the consistent write-back unit.

[0073] Grid-level indicators directly guide the execution of control actions such as humidification, ventilation, or supplemental lighting in simulated ecological cultivation by identifying the growth stage and confidence level of morel mushrooms within each grid cell. Specifically, the grid-level indicator set includes the stage label (e.g., primordia, young mushroom, or mature mushroom) and its corresponding confidence level for each grid cell, reflecting the main growth state and reliability of the cap within that grid. Control devices (such as sprayers, fans, or supplemental lighting) receive the grid-level indicators and perform corresponding environmental adjustments according to preset stage control rules. For example, if the grid-level indicator of a certain bed cell shows a stage label of primordia with high confidence, it indicates that the cap is in the early growth stage and requires a high humidity environment. The control action triggers the sprayer to increase local humidification, keeping the bed surface moist to promote primordia development. If the stage label is mature mushroom with high confidence, it means the cap is nearing harvest, and humidity needs to be reduced to prevent over-wetting. The control action activates the fan to enhance ventilation while simultaneously turning off the sprayer. If the stage label is young mushroom with high confidence, it indicates the cap is in a rapid growth stage and may require supplemental lighting to promote morphological development. The control action activates the supplemental lighting to provide suitable light intensity. These control actions precisely match the growth needs of morel mushrooms through the stage information of the grid-level indicator, ensuring that environmental parameters are consistent with the cap development stage, thereby improving cultivation efficiency and yield.

[0074] The stage calibration unit extracts the growth dynamics and humidity characteristics of the cap from the instance trajectory set and the original image sequence through instance trajectory reception, morphological evolution calculation, neighborhood wet mark extraction, stage label generation, and grid-level indicator aggregation, generating accurate stage labels and grid-level indicators. The joint analysis of morphological evolution and wet mark features effectively distinguishes between primordia, young mushrooms, and mature mushrooms, while grid-level aggregation ensures the stability of control commands. The grid-level indicator set directly supports the consistent write-back unit for control triggering and parameter optimization, improving the closed-loop control effect of biomimetic cultivation.

[0075] After the control action is executed, environmental changes may cause inconsistencies between image features and stage labels, requiring optimization of preceding processing parameters through a feedback mechanism. The consistency write-back unit receives new image sequences and grid-level indicator sets, extracts consistency signals, and calibrates preceding unit parameters to ensure continuous accuracy in identification and control.

[0076] The following details the specific processing logic of the consistent write-back unit: 5.1: New image reception.

[0077] The system receives new image sequences captured by the inspection camera after a control action is triggered. These sequences contain pixel coordinates and intensity values ​​for the red, green, and blue channels of each frame, with the intensity values ​​normalized to the range of zero to one. Simultaneously, it receives a set of grid-level indicators generated by the stage calibration unit. This set includes the grid-level stage label and confidence level for each grid cell in the bed, reflecting the growth stage and reliability of the cap. The new image sequences reflect the bed surface condition after control actions (such as humidification, ventilation, or supplemental lighting), and the set of grid-level indicators provides a reference for the current growth stage.

[0078] 5.2: Short-term incremental extraction of pore direction.

[0079] Based on the new image sequence, short-term increments of pore orientation in each bed grid unit were extracted to reflect changes in cap texture direction after modulation. For each frame, within the pixel region of the bed grid unit, the pore orientation field was calculated: a fixed-size window centered on each pixel was taken, and the changes in intensity of the red, green, and blue channels in the horizontal and vertical directions were calculated to determine the intensity gradient direction angle. The direction angle with the highest frequency within the window was selected as the pore orientation field value. The short-term increment of pore orientation is the absolute difference between the average value of the pore orientation field in the current frame and the average value of the pore orientation field in the previous frame. If the previous frame is missing, the most recent valid frame is used. The set of short-term increments of pore orientation reflects the dynamic changes in cap texture after modulation.

[0080] 5.3: Wet trace shrinkage extraction.

[0081] Based on the new image sequence, wettation shrinkage features of each bed grid cell were extracted to reflect changes in humidity distribution after regulation. For each image frame, wettation regions were identified within the pixel area of ​​the bed grid cell. A wettation region is a set of pixels with an intensity below a wettation intensity threshold and a blue channel intensity higher than the red and green channels. The wettation shrinkage feature is the difference between the proportion of wettation region pixels in the current frame to the total number of pixels in the grid cell and the corresponding proportion in the previous frame. If the previous frame is missing, the most recent valid frame is used. The set of wettation shrinkage features reflects the trend of humidity changes after regulation.

[0082] 5.4: Consistency signal verification.

[0083] By combining the short-term increment set of pore orientation and the wet mark shrinkage feature set, the consistency with the grid-level indicator set is verified, and a consistency signal is generated. For each bed grid cell, the expected range of short-term increment of pore orientation and wet mark shrinkage is set according to the grid-level stage label. For example, the expected increment is small and the shrinkage is slow in the primordium stage, while the expected increment is large and the shrinkage is fast in the mature mushroom stage. The consistency signal is defined as follows: if both the short-term increment of pore orientation and the wet mark shrinkage feature are within the expected range and the confidence level is higher than the confidence level threshold, it is marked as consistent; otherwise, it is marked as inconsistent. The consistency signal set indicates the degree of matching between the adjusted features and the stage label. The quantification method of the confidence level threshold is determined by statistically analyzing the distribution of confidence levels in the grid-level indicator set. For example, the average confidence level of samples at different growth stages is calculated, and a fixed proportion of the average value is selected as the threshold to ensure the reliability of the stage label.

[0084] 5.5: Parameter write-back calibration.

[0085] Based on the consistency signal set, the parameters of the preceding cells are calibrated for the grid cells marked as inconsistent. This includes the candidate seed generation rules for the grid seeding cell, the instance segmentation boundary parameters and identity preservation rules for the instance tracking cell, and the stage label rules for the stage calibration cell. Calibration is achieved by adjusting the thresholds of each cell: if the short-term increment of the pore orientation deviates from the expectation, the orientation consistency threshold and fine-grained texture density threshold for candidate seed generation are adjusted; if the wet mark shrinkage deviates from the expectation, the intensity difference threshold for instance segmentation, the distance threshold for identity preservation, and the area change rate threshold (including the first and second area thresholds), aspect ratio change rate threshold, and wet mark threshold for the stage label are adjusted. The adjusted parameters are referenced using the grid coordinates and written back to the preceding cells to ensure that the processing adapts to environmental changes.

[0086] The following detailed example illustrates the calibration method for a simulated morel cultivation scenario, ensuring that the calibration adapts to environmental changes after regulation and improving the accuracy of identification and regulation. The calibration is based on deviation analysis of short-term increments in the pore orientation and wett mark shrinkage characteristics, combined with sample statistics to adjust the threshold, and uses the center coordinates of the bed grid cells for reference.

[0087] Example: How to calibrate the threshold of each unit.

[0088] Suppose a certain grid cell in a bed is marked as inconsistent by the consistency signal set, indicating that the short-term increment of the pore orientation or the wett mark shrinkage characteristics do not match the expectations of the grid-level stage label. For example, if the grid-level stage label is primordium, the expected short-term increment of the pore orientation is small (because the cap texture is stable in the primordium stage) and the wett mark shrinkage is slow (due to the need for high humidity), but the actual extracted short-term increment of the pore orientation is large (the texture direction changes drastically) and the wett mark shrinkage is fast (the wett mark area decreases rapidly). The following is a specific calibration of the threshold for each cell: 1. Calibration of candidate seed generation rules for grid seeding units: Candidate seed generation depends on the orientation consistency threshold and the fine-grained texture density threshold. If the short-term increment of the pore orientation is too large, it indicates that the candidate seed may contain non-cap regions (such as mud or dead leaves), and the orientation consistency threshold needs to be tightened to more strictly screen cap textures. The calibration method is as follows: In the new image sequence, analyze the pore orientation field of the bed grid unit, statistically analyze the orientation distribution of the cap region, and select the high percentile value of the orientation difference as the new orientation consistency threshold to replace the original threshold; if the fine-grained texture density is lower than expected, it indicates that the cap features have not been fully extracted, and the fine-grained texture density threshold needs to be reduced, selecting the median of the density distribution of the cap region as the new threshold. This adjustment ensures that the candidate seed more accurately locates the cap and reduces background interference.

[0089] 2. Instance segmentation boundary parameter calibration of the instance tracking unit: The instance segmentation boundary parameters include an intensity difference threshold and an orientation consistency threshold. If the short-term increment of the pore direction is too large, it indicates that the segmentation boundary may contain non-cap pixels, and the intensity difference threshold needs to be increased to tighten the growth range of the region. The calibration method is as follows: In the new image sequence, for the cap region of the bed grid unit, the intensity difference distribution between the cap and background pixels is statistically analyzed, and the high percentile value is selected as the new intensity difference threshold; if the orientation consistency threshold causes the boundary to stick together, the distribution of the pore direction in the cap region is statistically analyzed, and the orientation difference value with higher concentration is selected as the new threshold. This adjustment improves the accuracy of instance segmentation boundaries and reduces boundary sticking or breakage.

[0090] 3. Instance tracking unit identity preservation rule calibration: The identity preservation rule relies on a center coordinate distance threshold. If the wettrace shrinkage feature shows a rapid decrease, it indicates that changes in viewpoint or occlusion may cause a shift in the center coordinates of the instance trajectory, requiring a relaxation of the distance threshold to maintain identity continuity. The calibration method is as follows: in a new image sequence, analyze the distribution of center coordinate shifts of instance trajectories within the bed grid cell, and select the high percentile value as the new distance threshold. This adjustment ensures that instance trajectories maintain identity continuity under changes in viewpoint or occlusion, reducing identity drift.

[0091] 4. Stage label rule calibration for stage calibration units: The stage labeling rules involve thresholds for area change rate, aspect ratio change rate, and wettage threshold. If wettage shrinkage characteristics show a rapid decrease while the grid-level stage label is primordium (expected to shrink slowly), it indicates that the wettage threshold is too high and needs to be lowered to adapt to actual humidity changes. The calibration method is as follows: In the new image sequence, the distribution of wettage area proportion within the bed grid unit is statistically analyzed, and the lower percentile value of the primordium stage sample is selected as the new wettage threshold; if the area change rate or aspect ratio change rate does not match the primordium stage, the distribution of morphological evolution characteristics of the cap sample is statistically analyzed, and the median of the corresponding stage is selected as the new area change rate threshold or aspect ratio change rate threshold. This adjustment ensures that the stage label accurately reflects the cap growth stage.

[0092] Calibration is performed on the new image sequence. For bed grid cells marked as inconsistent, deviations in short-term increments of porosity orientation and wett mark shrinkage characteristics are analyzed, and new thresholds are determined based on sample statistics. The calibration parameter set includes orientation consistency threshold, fine-grained texture density threshold, intensity difference threshold, center coordinate distance threshold, area change rate threshold, aspect ratio change rate threshold, and wett mark threshold. The center coordinate references of the bed grid cells are used and written back to the grid seeding unit, instance tracking unit, and stage calibration unit to update their respective processing rules. This calibration method dynamically adjusts the thresholds to adapt to changes in the bed surface environment after control actions (such as decreased humidity or texture changes), ensuring the accuracy of candidate seed location, instance segmentation, identity tracking, and stage discrimination, and maintaining the stability of closed-loop control.

[0093] In this example, for the inconsistency signals of the bed grid cells (large short-term increment of pitting direction and rapid shrinkage of wet marks in the primordium stage), the calibration parameter set was optimized by increasing the orientation consistency threshold, decreasing the fine texture density threshold, tightening the intensity difference threshold, relaxing the center coordinate distance threshold, decreasing the wet mark threshold, and adjusting the area change rate threshold. This optimized the processing of the preceding cells, making the identification results consistent with the regulated environment and improving the regulation accuracy and stability of morel mushroom biomimetic cultivation.

[0094] The consistency write-back unit, through new image reception, short-term incremental extraction of pore orientation, extraction of wet streak shrinkage, consistency signal verification, and parameter write-back calibration, verifies the matching degree between features and stage labels based on new image sequences and grid-level indicator sets, and calibrates the parameters of preceding units. Short-term incremental extraction of pore orientation and wet streak shrinkage characteristics reflect the control effect; consistency verification and parameter write-back optimize the identification link; and calibrating the parameter set ensures the accuracy of grid seeding, instance tracking, and stage calibration, maintaining the stability of the closed-loop control of biomimetic cultivation.

[0095] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0096] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0097] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0098] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent control system for the simulated ecological cultivation of morel mushrooms based on image recognition, characterized in that, include: Imaging and normalization unit: Acquires multi-view sequences of the biomimetic bed surface and generates reflection normalized frames; Mesh seeding unit: Align the reflection normalized frame geometry and construct the bed mesh. Generate candidate seeds in each mesh only in the intersection area of ​​the pore orientation field and the fine-grained texture density map, and bind the mesh coordinates. Instance tracking unit: Starting from the candidate seed, it performs instance segmentation to obtain the instance segmentation boundary and establishes the instance trajectory. It maintains the continuity of identity under the grid coordinate constraint according to the identity preservation rule and fills in the missing segments when there is occlusion and viewpoint change. Stage calibration unit: Calculates morphological evolution along the instance trajectory and generates stage labels based on stage label rules in conjunction with neighboring wet marks. The stage labels are aggregated into grid-level indicators and output trigger guidelines. Consistent write-back unit: In the new image acquired after triggering, the short-term increment of the porosity direction and the shrinkage of the wet mark are extracted as consistency signals. If the consistency signal is inconsistent with the stage label, the candidate seed generation rule, instance segmentation boundary parameters, identity preservation rule, and stage label rule are written back synchronously, and the grid coordinate reference is used during the write-back.

2. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 1, characterized in that: The imaging and normalization unit acquires multi-view image sequences of the biomimetic bed surface, generates an original image sequence, and sequentially performs reflection suppression on each frame to reduce the intensity of the water film highlight area while preserving texture details, generating a suppressed reflection image sequence. Shadow correction is then performed on the suppressed reflection image sequence to adjust local brightness and eliminate the influence of striped shadows, generating a shadow-corrected image sequence. Geometric alignment is performed on the shadow-corrected image sequence, and scale and perspective drift are corrected based on the reference frame through feature point matching, generating an aligned image sequence. Finally, the temporal median is calculated for the aligned image sequence based on pixel position and channel, generating a reflection-normalized frame with uniform brightness and clear texture.

3. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 2, characterized in that: The grid seed-taking unit receives the reflection normalized frame, aligns it to the bed surface reference coordinate system to generate an aligned reflection frame, constructs a bed grid to divide the bed surface into regular grid units, extracts the pore orientation field to obtain the cap texture direction features, generates a fine-grained texture density map to highlight the density of cap details, generates candidate seeds based on the intersection area of ​​the pore orientation field and the fine-grained texture density map in each grid unit and binds them to the grid coordinates, and outputs a set of candidate seeds for use by the instance tracking unit.

4. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 3, characterized in that: The instance tracking unit receives a set of candidate seeds and the original image sequence, initializes an initial instance set based on the candidate seeds, performs instance segmentation on each frame of the image to generate instance segmentation boundaries, constructs instance trajectories to record the temporal changes of the cap, applies bed grid center coordinate constraints to maintain the continuity of instance identity, fills in the boundary missing caused by occlusion or viewpoint changes, and outputs a set of instance trajectories.

5. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 4, characterized in that: The stage calibration unit receives the set of instance trajectories and the original image sequence, calculates the morphological evolution features of the instance trajectories to extract the cap growth dynamics, extracts the neighborhood wet mark features of the bed grid unit to reflect the humidity distribution, combines the morphological evolution features and neighborhood wet mark features to generate stage labels to identify the cap growth stage, aggregates the stage labels as grid-level indicators, and outputs the set of grid-level indicators for use by the consistent write-back unit.

6. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 3, characterized in that: The acquisition of the aperture orientation field is based on local texture analysis of aligned reflection frames: for each pixel position, a fixed-size window is selected with the pixel as the center, and the variation of the intensity of the red, green and blue channels in the horizontal and vertical directions within the window is calculated to determine the intensity gradient direction angle of each pixel; by statistically analyzing the distribution of all direction angles within the window, the direction angle with the highest frequency of occurrence is selected as the aperture orientation field value of the corresponding pixel.

7. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 3, characterized in that: The fine-grained texture density map is obtained by extracting the frequency domain features of the aligned reflection frames: for each pixel position, a small window is selected with the pixel as the center, and the pixel intensity of the red, green and blue channels in the window is transformed in the frequency domain to extract the high-frequency components and accumulate their square amplitude as the texture detail intensity; the intensity value is normalized to the range of zero to one.

8. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 5, characterized in that: The acquisition of neighborhood wet stain features is based on pixel analysis of the original image sequence within the bed grid cell: For each frame of image, wet stain regions are identified within the pixel area of ​​each bed grid cell. A wet stain region is defined as a set of pixels with pixel intensity below the wet stain intensity threshold and blue channel intensity higher than red and green channel intensity. The wet stain intensity threshold is determined through sample statistics. The proportion of the number of pixels in the wet stain region to the total number of pixel regions in the grid cell is calculated as the neighborhood wet stain feature.

9. The intelligent control system for morel mushroom biomimetic cultivation based on image recognition according to claim 5, characterized in that: The consistent write-back unit receives new image sequences and a set of grid-level indicators, extracts short-term increments of porosity orientation and wett shrinkage features of bed grid units, verifies consistency with grid-level stage labels, calibrates candidate seed generation rules of the grid seeding unit, instance segmentation boundary parameters and identity preservation rules of the instance tracking unit, and stage label rules of the stage calibration unit for inconsistent bed grid units, outputs a set of calibration parameters and references the center coordinates of the bed grid units.

Images