Blueberry fruit detection method and system based on image recognition

Abstract

This invention relates to the field of image recognition technology, specifically to a method and system for blueberry fruit detection based on image recognition. The method involves imaging and monitoring blueberry fruits using cross-polarization imaging technology to obtain polarized images and standard images of blueberries. The standard image is used to identify the developmental stage of the blueberry. Based on the developmental stage, an image segmentation strategy is selected to adaptively segment the polarized image, obtaining a bloom mask. The polarized image, standard image, and bloom mask are identified to obtain bloom features. Time-stamped bloom features are acquired to obtain bloom temporal features. The evolution of bloom features and the synergistic relationship between the bloom features and the fruit peel background color are analyzed and compared with standard temporal features to classify the fruit into normal and abnormal development types. This invention outputs fruit detection information through the identification of bloom temporal features of normal and abnormal development types.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a method and system for detecting blueberry fruits based on image recognition. Background Technology

[0002] As a high-value-added economic crop, the quality of blueberries, especially their ripeness and health, directly determines their commercial value. In traditional blueberry cultivation and management, the judgment of fruit ripeness and the screening for pests and diseases mainly rely on manual visual inspection and sampling. This method is not only highly subjective, inefficient, and labor-intensive, but also cannot achieve large-scale, continuous, and precise monitoring of the entire crop growth process.

[0003] With the development of computer vision technology, automated inspection of agricultural products using image recognition methods has become a research hotspot. Most existing solutions are based on standard RGB color images, locating fruits through target detection algorithms, extracting their contours using image segmentation techniques, and further extracting macroscopic features such as color, size, and shape. However, these methods face significant technical bottlenecks when applied to blueberry inspection. First, the unique "bloom" (white powder) on the surface of blueberries has complex reflective properties, producing severe highlights and shadows under different lighting conditions, greatly interfering with the accuracy of traditional image segmentation algorithms. Second, the presence of the bloom obscures the true base color of the fruit skin, significantly reducing the reliability of methods that rely solely on apparent color to determine ripeness.

[0004] To address this, a blueberry fruit detection method and system based on image recognition is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a blueberry fruit detection method and system based on image recognition. The method uses cross-polarization imaging technology to image and monitor blueberry fruits, obtaining blueberry polarization images and blueberry standard images. Based on the blueberry polarization images and blueberry standard images, the method detects blueberries and outputs detection information.

[0006] To achieve the above objectives, the present invention provides a blueberry fruit detection method based on image recognition, comprising:

[0007] Blueberry fruits were imaged and monitored using cross-polarization imaging technology to obtain polarized images and standard images of blueberries.

[0008] A blueberry development stage recognition model is constructed to identify the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, an image segmentation strategy is selected to adaptively segment the blueberry polarization image and identify the white bloom mask;

[0009] A fruit bloom feature quantification model was constructed to identify blueberry polarized images, standard blueberry images, and white bloom masks to obtain white bloom features, specifically including fruit bloom coverage, fruit bloom thickness, fruit bloom uniformity, and fruit skin background color quantification values.

[0010] Continuously acquire time-tagged white bloom features to obtain bloom time-series features; construct a blueberry dynamic development model to identify the bloom time-series features, analyze the synergistic relationship between the evolution of bloom features and the background color of the peel, and compare and analyze with standard time-series features to classify them into normal development type and abnormal development type.

[0011] For fruit bloom time-series characteristics of abnormal development types, the abnormality type is identified and an early warning is issued; for fruit bloom time-series characteristics of normal development types, predictions are made in conjunction with environmental prediction data, the maturity date is determined, and the result is output.

[0012] The steps for imaging and monitoring blueberry fruit based on cross-polarization imaging technology include:

[0013] An active light source system is used, which integrates a first linear polarizer to emit polarized light onto the blueberry fruit; at the imaging end, a high-resolution image sensor is used, and a second linear polarizer is deployed in front of the lens as an analyzer.

[0014] The polarization direction of the analyzer is controlled to be parallel to the polarization direction of the first linear polarizer, and a standard blueberry image containing specular reflection light from the fruit surface is acquired to identify the outline, size and apparent color of the blueberry fruit.

[0015] The polarization direction of the analyzer is rotated by 90 degrees so that it is orthogonal to the polarization direction of the first linear polarizer. Under cross-polarization, light that has been depolarized by multiple scatterings within the white bloom is collected to obtain a blueberry polarization image that highlights the spatial distribution and density of the white bloom.

[0016] The steps to build a blueberry stage identification model and determine the blueberry development stage include:

[0017] Standard images covering all stages of blueberries from immature to fully ripe were collected and labeled according to the outline, size and apparent color of the blueberries to create a training dataset with three label types: young fruit stage, color change stage and fully ripe stage.

[0018] A deep learning-based convolutional neural network architecture is used as the foundation of the blueberry stage recognition model; supervised learning training is performed on the training dataset to enable the model to learn to extract multi-dimensional features of color, texture and shape from standard images;

[0019] During the identification phase, the collected standard blueberry images are input into the trained blueberry stage identification model. Through forward propagation calculation, the probability of each developmental stage is output, and the stage with the highest probability value is determined as the blueberry developmental stage.

[0020] The process of adaptive segmentation of blueberry polarization images based on image segmentation strategies selected according to blueberry development stages includes:

[0021] For blueberry polarized images in the young fruit stage of blueberry development, image segmentation is performed based on a preset severity threshold;

[0022] For blueberry polarized images in the color-changing stage of blueberry development, a local adaptive thresholding algorithm is used for image segmentation. The local adaptive thresholding algorithm dynamically calculates the local threshold based on the global distribution within the neighborhood of each pixel.

[0023] For polarized images of blueberries at the fully ripe stage of development, the Otsu adaptive thresholding algorithm was used for image segmentation.

[0024] The characteristics of the white bloom fruit powder include the fruit powder coverage rate, fruit powder coverage thickness, fruit powder uniformity, and quantitative value of the fruit peel background color;

[0025] The fruit bloom coverage rate is obtained by applying a white bloom mask to the total pixel area of ​​the fruit and calculating the ratio of the number of pixels in the mask to the total number of pixels in the fruit.

[0026] The fruit bloom coverage thickness is determined by extracting the pixel grayscale values ​​of the corresponding region of the white bloom mask in the blueberry polarization image, and calculating the average value of the pixel grayscale values ​​as a quantitative indicator of the fruit bloom coverage thickness.

[0027] The uniformity of the fruit bloom is determined by calculating the standard deviation of pixel gray values ​​within the white bloom mask area in a blueberry polarized image to assess the dispersion of the fruit bloom distribution, which is then used as the uniformity of the fruit bloom.

[0028] The quantization value of the fruit peel background color is obtained by locating the pixel area corresponding to the white bloom mask in the blueberry standard image and removing it to obtain the exposed fruit peel area; the chromaticity value of the exposed fruit peel area in the color space is extracted and calculated, and quantized into the fruit peel background color quantization value.

[0029] The steps for constructing a dynamic development model of blueberries and classifying developmental types include:

[0030] A recurrent neural network capable of processing time-series data was adopted as the core architecture of the blueberry dynamic development model.

[0031] Using historical monitoring data of blueberry bloom time-series feature samples marked as normally developing, the blueberry dynamic development model was trained to learn and encode the inherent laws and standard development patterns of bloom coverage, bloom thickness, bloom uniformity and peel background color quantification values ​​co-evolving over time.

[0032] During model recognition, the dynamic evolution trend of fruit powder time sequence characteristics is analyzed through the blueberry dynamic development model, and similarity matching and deviation calculation are performed with the internally learned standard development pattern to obtain the development deviation index; then, based on the preset deviation threshold, the normal development type and abnormal development type are determined.

[0033] Furthermore, at time point t, the time-series features of Apple fans can be represented as a four-dimensional vector F. t :

[0034] F t =[C r (t), T c (t), U f (t), C p (t)];

[0035] Among them, C r (t) represents the fruit powder coverage at time point t; T c (t) represents the thickness of the bloom coverage at time t; U f (t) represents the uniformity of fruit powder at time point t; C p (t) represents the quantification value of the fruit peel background color at time point t.

[0036] Once the fruit powder time sequence characteristics are determined to be an abnormal development type, the rule-based and pattern recognition-based diagnostic module will be activated.

[0037] The diagnostic module analyzes deviation patterns in fruit pollen time-series characteristics that deviate from the standard developmental pattern, and identifies abnormality types based on these deviation patterns.

[0038] The process for predicting the maturity date of the fruit powder timing characteristics of the normal development type includes:

[0039] A multivariate regression prediction model integrating pollen time-series characteristics and environmental factors was constructed. The multivariate regression prediction model uses the pollen time-series characteristics of normal development type as the core input and integrates environmental prediction data.

[0040] During the training phase of the multiple regression prediction model, historical maturity data is used to learn the time period required for a normally developing blueberry to evolve to its final mature state, as well as the mapping relationship between the changes in white bloom characteristics and environmental data during the time period.

[0041] In the prediction phase of the multiple regression prediction model, the current fruit pollen time-series characteristics are received, and combined with the acquired environmental prediction data, the future evolution data of the fruit pollen time-series characteristics are dynamically deduced; the time required for the future evolution data to reach the preset maturity threshold is calculated, and the maturity date is determined.

[0042] This invention also proposes an image recognition-based blueberry fruit detection system, comprising:

[0043] The image acquisition module uses cross-polarization imaging technology to monitor blueberry fruits and obtain blueberry polarization images and standard blueberry images.

[0044] The fruit bloom recognition module constructs a blueberry stage recognition model to recognize the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, it selects an image segmentation strategy to adaptively segment the blueberry polarization image and recognizes the white bloom mask.

[0045] The feature quantization module constructs a fruit bloom feature quantification model to identify blueberry polarized images, standard blueberry images, and white bloom masks, and obtain white bloom features, specifically including fruit bloom coverage, fruit bloom coverage thickness, fruit bloom uniformity, and fruit skin background color quantification values.

[0046] The type classification module continuously acquires the white bloom fruit powder features with time tags to obtain the fruit powder temporal features; constructs a blueberry dynamic development model to identify the fruit powder temporal features, analyzes the synergistic relationship between the evolution of fruit powder features and the background color of the fruit peel, and compares and analyzes them with standard temporal features to classify them into normal development type and abnormal development type.

[0047] The output module determines the abnormality type and outputs a warning for fruit pollen time-series characteristics of abnormal development types; for fruit pollen time-series characteristics of normal development types, it combines environmental prediction data to predict the maturity date and output it.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. This invention uses active light source and cross-polarization technology to physically filter out interference in the natural environment, which greatly enhances the signal of white bloom in the output blueberry polarized image, while greatly suppressing the background and interference areas; ensuring the accuracy and repeatability of the white bloom features quantified from the image, laying a solid foundation for the accuracy and robustness of the entire detection and prediction system.

[0050] 2. This invention introduces a deep learning-based blueberry stage recognition model to accurately determine the development stage of the blueberry based on its outline, size, and apparent color. This upgrades the entire detection system from a passive and blind processing flow to an active and context-aware intelligent system. It also enables intelligent guidance of subsequent image processing flows, providing a technical foundation for subsequent adaptive segmentation strategies.

[0051] 3. This invention uses the fruit development stage as a decision input and dynamically switches intelligently between multiple optimal algorithms to ensure that image segmentation is performed in the most suitable way. This fundamentally guarantees the authenticity and reliability of core features such as fruit bloom coverage, thickness, and uniformity in subsequent calculations, and directly and significantly achieves precision control and robustness in the extraction of white bloom mask.

[0052] 4. The blueberry dynamic development model of the present invention endows the detection system with process-oriented and predictive diagnostic capabilities; by analyzing the dynamic evolution trend of fruit bloom time-series characteristics through the blueberry dynamic development model, and performing similarity matching and deviation calculation with the internally learned standard development model, a development deviation index is obtained; then, based on the preset deviation threshold, the normal development type and abnormal development type are determined; it can detect abnormalities such as weak growth and development stagnation at an early stage, significantly improving the precision and foresight of blueberry fruit management. Attached Figure Description

[0053] Figure 1 This is a schematic flowchart of a blueberry fruit detection method based on image recognition according to the present invention.

[0054] Figure 2 This is a schematic diagram of the logic for determining the development type of blueberry fruit according to the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of a blueberry fruit detection system based on image recognition according to the present invention;

[0056] Figure 4 This is a schematic diagram of the blueberry grayscale image of the present invention. Detailed Implementation

[0057] 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.

[0058] Example 1:

[0059] This invention proposes a blueberry fruit detection method based on image recognition, the process of which is as follows: Figure 1 As shown, it includes:

[0060] Blueberry fruits were imaged and monitored using cross-polarization imaging technology, resulting in polarized blueberry images and standard blueberry images.

[0061] The steps for imaging and monitoring blueberry fruit based on cross-polarization imaging technology include:

[0062] An active light source system is used, which integrates a first linear polarizer to emit polarized light onto the blueberry fruit; at the imaging end, a high-resolution image sensor is used, and a second linear polarizer is deployed in front of the lens as an analyzer.

[0063] The polarization direction of the analyzer is controlled to be parallel to the polarization direction of the first linear polarizer, and a standard blueberry image containing specular reflection light from the fruit surface is acquired to identify the outline, size and apparent color of the blueberry fruit.

[0064] The polarization direction of the analyzer is rotated by 90 degrees so that it is orthogonal to the polarization direction of the first linear polarizer. Under cross-polarization, light that has been depolarized by multiple scatterings within the white bloom is collected to obtain a blueberry polarization image that highlights the spatial distribution and density of the white bloom.

[0065] Preferably, an active light source system refers to a device that is controlled by the system and has its own lighting, rather than a passive system that relies on ambient natural light. In this embodiment, it is a ring-shaped LED light source array.

[0066] Linear polarizer: An optical element that allows light waves with a specific vibration direction to pass through. The first linear polarizer (polarizer) is used to generate polarized light in a specific direction, and the second linear polarizer (analyzer) is used to analyze the polarization information in the reflected light.

[0067] Cross-polarization state: refers to the configuration where the polarizer and analyzer have polarization directions that are perpendicular to each other (forming a 90° angle).

[0068] The smooth skin of blueberries primarily produces specular reflection of polarized light, with the reflected light largely maintaining its original polarization direction. However, in cross-polarization, this polarized light, maintaining its polarization direction, is physically filtered out by an analyzer orthogonal to it, causing the smooth skin area to appear very dark in a blueberry polarized image. Conversely, the bloom covering the skin is composed of countless microscopic waxy crystals with an extremely rough surface. Light entering the bloom undergoes multiple scattering processes, completely disrupting the polarization direction of the reflected light (depolarization). This depolarized light contains components with various vibrational directions, so some of it always passes through the analyzer, making the bloom area appear bright in a blueberry polarized image.

[0069] In natural environments, uneven lighting, leaf movement, and water droplets on the fruit surface can all generate strong interfering reflections. These reflections severely interfere with traditional RGB image recognition technology, leading to deviations in image segmentation and feature extraction. This invention uses an active light source and cross-polarization technology to physically filter out interference from the natural environment. This results in a significant enhancement of the white bloom signal in the output blueberry polarized image, while background and interference areas are greatly suppressed. This ensures the accuracy and repeatability of the white bloom features quantified from the image, laying the foundation for the accuracy and robustness of the entire detection and prediction system.

[0070] A blueberry development stage recognition model is constructed to identify the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, an image segmentation strategy is selected to adaptively segment the blueberry polarization image and identify the white bloom mask.

[0071] The steps to build a blueberry stage identification model and determine the blueberry development stage include:

[0072] Standard images covering all stages of blueberries from immature to fully ripe were collected and labeled according to the outline, size and apparent color of the blueberries to create a training dataset with three label types: young fruit stage, color change stage and fully ripe stage.

[0073] A deep learning-based convolutional neural network architecture is used as the foundation of the blueberry stage recognition model; supervised learning training is performed on the training dataset to enable the model to learn to extract multi-dimensional features of color, texture and shape from standard images;

[0074] During the identification phase, the collected standard blueberry images are input into the trained blueberry stage identification model. Through forward propagation identification, the probability of each developmental stage is output, and the stage with the highest probability value is determined as the blueberry developmental stage.

[0075] Forward propagation is the process of analyzing blueberry images based on a trained recognition model to determine their developmental stage; first, standard blueberry images are input into the model in the form of a matrix of digital pixels.

[0076] In the convolutional layers of the model, basic visual features such as edges, colors (e.g., the green of young fruit or the dark blue of ripe fruit), and textures are extracted. These basic visual features are then combined into advanced features in deeper network layers, precisely corresponding to different developmental stages.

[0077] Then, these refined high-level features are fed into a fully connected layer for comprehensive voting, and different weights are assigned according to their importance. Finally, the output layer calculates these weights into three distinct probability values, corresponding to the "young fruit stage", "color-changing stage", and "full ripening stage", respectively.

[0078] For example, given an input blueberry image, the output layer will provide three probability values, which sum to 1: the probability of being in the "young fruit stage" is 0.05; the probability of being in the "color-changing stage" is 0.15; and the probability of being in the "fully ripe stage" is 0.80. Therefore, it is determined that the image belongs to the fully ripe stage.

[0079] For example, the data collection and labeling process is as follows: During a complete blueberry growing season, blueberry fruits are photographed regularly to obtain thousands of standard images. Based on the outline, size, and apparent color of the blueberry fruits, each fruit in the image is labeled with its stage; specifically, this includes "young fruit stage" (green, small, firm), "color change stage" (pink / purple, nearing maturity, beginning to soften), and "fully ripe stage" (deep blue, with a uniform white bloom on the surface); ultimately forming a training dataset containing the labeled fruits.

[0080] For example, the model building and training process is as follows: a lightweight CNN architecture, such as MobileNetV3, is selected to adapt to efficient operation on embedded devices; the labeled dataset is input into the network for supervised learning training using the PyTorch deep learning framework; and the network weights are continuously adjusted through the backpropagation algorithm and the Adam optimizer to minimize the cross-entropy loss between the prediction stage and the true label.

[0081] CNN architectures can automatically learn richer and more robust visual features from pixel data than simple average color. For example, it can learn not only that "green" represents young fruit, but also the smooth texture and small outline characteristic of young fruit; similarly, it can associate "low-saturation blue" with full ripeness. In this way, it can make more accurate and reliable judgments about the developmental stage of fruit than simple color thresholding methods.

[0082] This invention introduces a deep learning-based blueberry stage recognition model to accurately determine the development stage of blueberries based on their outline, size, and apparent color. This upgrades the entire detection system from a passive, blind processing flow to an active, context-aware intelligent system. It also enables intelligent guidance of subsequent image processing flows, providing a technical foundation for subsequent adaptive segmentation strategies.

[0083] The process of adaptive segmentation of blueberry polarization images based on image segmentation strategies selected according to blueberry development stages includes:

[0084] For blueberry polarized images in the young fruit stage of blueberry development, image segmentation is performed based on a preset severity threshold;

[0085] For blueberry polarized images in the color-changing stage of blueberry development, a local adaptive thresholding algorithm is used for image segmentation. The local adaptive thresholding algorithm dynamically calculates the local threshold based on the global distribution within the neighborhood of each pixel.

[0086] For polarized images of blueberries at the fully ripe stage of development, the Otsu adaptive thresholding algorithm was used for image segmentation.

[0087] The principle of image segmentation strategy is to match the most suitable segmentation tool to images with different characteristics;

[0088] For young fruit: The severity threshold is a globally fixed threshold set for polarized images of young blueberries, used for image segmentation. At this stage, the fruit surface is smooth, and theoretically, there should be no or only trace amounts of waxy bloom; bright areas in the image are highly likely to be noise or artifacts. Therefore, a very high global threshold is applied as the severity threshold to segment the blueberry polarized image, identifying only extremely bright outliers and generating bloom masks, while ignoring all normal, darker skin areas. The severity threshold is determined by performing grayscale histogram analysis on at least 1000 polarized images of young blueberries from different batches, selecting grayscale values ​​that can exclude more than 90% of skin pixels; for example, the severity threshold can be set to 245 (maximum 255); each pixel in the image is traversed, and its brightness value is checked; when the brightness value of a pixel is greater than 245, it is determined to be a potential bloom area; all pixels with brightness values ​​equal to or less than 245 are ignored or considered background.

[0089] For fruit in the ripening stage: At this stage, the white bloom begins to form, but its distribution is usually sparse, mottled, and uneven, and the brightness difference between it and the peel is not as significant as in the ripening stage. Therefore, the global thresholding method is abandoned in favor of a local adaptive thresholding algorithm. The local adaptive thresholding algorithm dynamically calculates the segmentation threshold based on the brightness distribution within the neighborhood of each pixel. This effectively extracts those faint early white bloom patches with brightness only slightly higher than the surrounding peel from the unevenly lit background, generating a white bloom mask. The size of this neighborhood can be set to 15x15 pixels.

[0090] For ripe fruit: At this stage, the white bloom is thick and evenly distributed, creating a very high contrast with the dark peel background, making its characteristics very obvious. In this scenario, the image's grayscale histogram will show a clear bimodal distribution, with one peak representing the dark background and the other representing the bright white bloom. The Otsu adaptive thresholding algorithm is a classic global adaptive thresholding algorithm that automatically finds an optimal value among all possible thresholds, dividing all pixels of the image into two classes (foreground and background) and maximizing the inter-class variance between these two classes. This algorithm performs particularly well when the image's grayscale histogram shows a clear bimodal distribution. Therefore, using the classic Otsu adaptive thresholding algorithm is the most efficient and robust choice, perfectly segmenting the thick white bloom layer as a complete and connected region to generate a white bloom mask.

[0091] This invention uses the fruit development stage as a decision input and dynamically switches intelligently between multiple optimal algorithms to ensure that image segmentation is performed in the most suitable way. This fundamentally guarantees the authenticity and reliability of core features such as fruit bloom coverage, thickness, and uniformity in subsequent calculations, and directly and significantly achieves precision control and robustness in the extraction of white bloom mask.

[0092] A fruit bloom feature quantification model was constructed to identify blueberry polarized images, standard blueberry images, and white bloom masks to obtain white bloom features, specifically including bloom coverage, bloom thickness, bloom uniformity, and fruit skin background color quantification values.

[0093] The fruit powder feature quantification model is implemented by coupling image processing algorithms and mathematical formulas, as follows:

[0094] The fruit bloom coverage rate is identified by applying a white bloom mask to the total pixel area of ​​the fruit and calculating the ratio of the number of pixels in the mask to the total number of pixels in the fruit. The fruit bloom coverage rate ranges from [0,1]. When the fruit bloom coverage rate is 0, it indicates that there is no white bloom at all. When the fruit bloom coverage rate is 1, it indicates that the fruit is completely covered by fruit bloom.

[0095] The formula for calculating the fruit powder coverage rate is as follows:

[0096]

[0097] Among them, C r Indicates fruit powder coverage; N mask N represents the number of pixels in the white frosting fruit powder mask; fruit The number of pixels within the total pixel area of ​​the fruit.

[0098] To identify the thickness of the bloom coverage, the pixel grayscale values ​​of the corresponding areas of the bloom mask in the blueberry polarized image are extracted, and the average value of the pixel grayscale values ​​is calculated as a quantitative indicator of the bloom coverage thickness. The bloom coverage thickness ranges from 0 to 255. When the value is 0, it means that the bloom area is completely non-reflective under the polarized image and appears pure black. When the value is 255, it means that the bloom area has the strongest reflectivity under the polarized image and appears pure white.

[0099] The formula for calculating the thickness of the fruit powder coating is:

[0100]

[0101] Among them, T c N represents the thickness of the pinkish coating, i.e., the average grayscale value of the pixels within the pinkish coating area; mask Indicates the number of pixels in the white frosting fruit powder mask; p i I(p) represents the pixels covering the area of ​​the white frosting mask; i ) represents pixel p i The grayscale value.

[0102] The uniformity of the bloom is calculated by measuring the standard deviation of the pixel grayscale values ​​within the bloom mask area in the blueberry polarized image to assess the dispersion of the bloom distribution, which is used as the bloom uniformity. The value of the bloom uniformity is greater than 0. When the value is 0, it indicates that the grayscale values ​​of all pixels are exactly the same, and the bloom is absolutely uniform. As the value of the bloom uniformity increases, it indicates that the bloom distribution is less uniform.

[0103] The formula for calculating the uniformity of the fruit powder is:

[0104]

[0105] Among them, U f Indicates the uniformity of the fruit bloom; I(p) i () represents the pixel grayscale value within the white frosting pink mask area;

[0106] The quantization value of the fruit peel background color is obtained by locating the pixel region corresponding to the white bloom mask in the blueberry standard image and removing it to obtain the exposed fruit peel region; the chromaticity value of the exposed fruit peel region in the color space is extracted and calculated, and quantized into the fruit peel background color quantization value. The value of the fruit peel background color quantization value ranges from 0 to 255, with a value of 0 corresponding to the start of the 8-bit color wheel and a value of 255 corresponding to the end of the 8-bit color wheel.

[0107] The formula for obtaining the quantitative value of the fruit peel background color is:

[0108]

[0109] Among them, C p N represents the quantitative value of the fruit peel's background color.peel q represents the total number of pixels in the area of ​​exposed fruit peel; j H(q) represents the pixels of the exposed fruit peel area; j ) represents pixel q j The hue value.

[0110] For example, the chromaticity value of the exposed fruit peel area in the HSV color space is extracted and calculated. Specifically, the average value of the H (hue) component of all pixels in the area is calculated, and the average value is used as the quantization value of the fruit peel background color.

[0111] The deletion operation excludes areas covered by fruit bloom from the total fruit area, thereby separating the bare fruit peel; specifically, the pixels of the bare fruit peel area must satisfy the condition of being "within the total fruit area" and "not within the area of ​​white bloom".

[0112] This invention characterizes the physical morphology of white bloom by measuring bloom coverage, bloom thickness, and bloom uniformity, and characterizes the biochemical state of the peel by measuring the background color of the peel. It constructs a comprehensive, objective, and information-rich digital archive for blueberry fruits, greatly improving the reliability of subsequent health diagnosis and the accuracy of maturity prediction.

[0113] The characteristics of white bloom with time tags are continuously acquired to obtain the time-series characteristics of the bloom; a blueberry dynamic development model is constructed to identify the time-series characteristics of the bloom, analyze the synergistic relationship between the evolution of the bloom characteristics and the background color of the peel, and compare and analyze them with the standard time-series characteristics to classify them into normal development type and abnormal development type.

[0114] The steps for constructing a dynamic development model of blueberries and classifying developmental types include:

[0115] A recurrent neural network capable of processing time series data is adopted as the core architecture of the blueberry dynamic development model. A recurrent neural network is a type of neural network specifically designed to process sequential data. Its internal recurrent structure gives it memory capabilities, enabling it to understand the contextual relationships in the time series.

[0116] Using historical monitoring data of blueberry bloom time-series feature samples marked as normally developing, a blueberry dynamic development model is trained to learn and encode the inherent laws and standard development patterns of bloom coverage, bloom thickness, bloom uniformity, and skin color quantification values ​​co-evolving over time. The standard development pattern refers to the standard template that the model internally encodes after learning from a large number of healthy samples, representing how various characteristics should evolve over time during the ideal blueberry growth process.

[0117] The historical monitoring data was obtained through continuous imaging and analysis of blueberry fruits from different batches and plants throughout their entire lifecycle, from the young fruit stage to full maturity. Each data record contains the following core information: a unique identifier for tracking individual fruits or clusters; a timestamp, accurate to the day or hour of collection; a feature vector, calculated by a feature quantification module, comprising four quantified dimensions: bloom coverage, bloom thickness, bloom uniformity, and skin color; and a status label, assigned by agricultural experts based on the fruit's final growth outcome, indicating normal or abnormal development. The data trajectory for normal development is shown in Table 1, and the data trajectory for abnormal development is shown in Table 2.

[0118] Table 1. Blueberry Normal Development Data Trajectory Table

[0119] Unique Identifier Timestamp Feature vector Status Label P08-C02-F01 2025-07-1509:00 [0.25,85,30.2,185] Normal development P08-C02-F01 2025-07-1809:00 [0.55,140,21.5,220] Normal development P08-C02-F01 2025-07-2109:00 [0.85,190,13.8,235] Normal development P08-C02-F01 2025-07-2409:00 [0.95,210,9.1,248] Normal development P08-C02-F01 2025-07-2409:00 [0.96,212,8.9,250] Normal development

[0120] In the initial stage, the bloom coverage is low (0.25), the thickness is not large (85 / 255), the uniformity is poor (standard deviation 30.2, indicating mottled), and the background color value is 185 (representing pink / purple, in the color-changing stage); in the mature stage, the bloom coverage is close to saturation (0.95), the thickness increases significantly (210 / 255), the uniformity is very good (standard deviation 9.1, indicating uniform bloom), and the background color value enters the dark blue / purple range (248); finally, all indicators tend to stabilize, indicating that it is fully ripe.

[0121] Table 2. Trajectory of Abnormal Blueberry Development Data

[0122] Unique Identifier Timestamp Feature vector Status Label P08-C05-F03 2025-07-1509:00 [0.30,90,28.5,190] Normal development P08-C05-F03 2025-07-1809:00 [0.60,155,19.8,225] abnormal development P08-C05-F03 2025-07-2109:00 [0.45,110,35.1,226] abnormal development P08-C05-F03 2025-07-2409:00 [0.42,95,38.4,228] abnormal development P08-C05-F03 2025-07-2409:00 [0.40,92,39.0,230] abnormal development

[0123] During the initial development stage, the fruit showed a normal development trend, with both the bloom coverage and thickness increasing. Subsequently, an abnormal mutation occurred, with a significant and irreversible decrease in bloom coverage and thickness, while the uniformity value increased dramatically (the standard deviation surged from 19.8 to 35.1, indicating severe damage to the bloom layer), consistent with the abnormal pattern of "premature degradation". The abnormality then persisted, and the fruit failed to recover normal growth, with the indicators continuing to deteriorate, confirming its abnormal development state.

[0124] During model recognition, the dynamic evolution trend of fruit bloom time-series characteristics is analyzed through the blueberry dynamic development model, and similarity matching and deviation calculation are performed with the internally learned standard development pattern to obtain the development deviation index; then, based on the preset deviation threshold, the normal development type and abnormal development type are determined; the development deviation index is a quantitative indicator used to describe the degree of difference between the real-time development trajectory of the test fruit and the corresponding standard development pattern in the model.

[0125] The data logic for determining normal and abnormal developmental types based on developmental deviation indices and deviation thresholds is as follows: Figure 2 As shown.

[0126] Specifically, similarity matching and deviation calculation can employ a dynamic time warping algorithm to calculate the distance between the temporal characteristic sequence of the fruit pollen to be tested and the standard developmental pattern sequence; this distance is the developmental deviation index. The deviation threshold is determined by receiver operating characteristic curve analysis of the developmental deviation indices of normal and abnormal samples to achieve optimal classification accuracy.

[0127] The blueberry dynamic development model of this invention endows the detection system with process-oriented and predictive diagnostic capabilities. By analyzing the dynamic evolution trend of fruit bloom time-series characteristics through the blueberry dynamic development model, and performing similarity matching and deviation calculation with the internally learned standard development patterns, a development deviation index is obtained. Then, based on the preset deviation threshold, the normal development type and abnormal development type are determined. This model can detect abnormalities such as weak growth and developmental arrest at an early stage, significantly improving the precision and foresight of blueberry fruit management.

[0128] Once the fruit powder time sequence characteristics are determined to be an abnormal development type, the rule-based and pattern recognition-based diagnostic module will be activated.

[0129] The diagnostic module analyzes deviation patterns in fruit pollen time-series characteristics that deviate from the standard developmental pattern, and identifies abnormality types based on these deviation patterns.

[0130] After the blueberry dynamic development model outputs the determination of the abnormal development type, the time-series characteristic data and deviation index of the fruit are passed to the diagnostic module, which makes a judgment according to the following rules:

[0131] Analysis of Deviation Patterns: The module first analyzes which feature dimensions the deviation occurs on, as well as the temporal characteristics of the deviation.

[0132] Apply diagnostic rules:

[0133] Rule A, Developmental Delay; This rule is designed to identify systemic slow or stunted growth of fruit due to persistent stress (such as malnutrition, water imbalance, root problems).

[0134] Judgment conditions:

[0135] Global developmental deviation: The developmental deviation index is consistently higher than a preset statistical significance threshold for five consecutive monitoring periods (e.g., five consecutive days). This statistical significance threshold is obtained by statistical analysis of the deviation index distribution of the healthy sample library, for example, by taking the upper bound of the 85% confidence interval.

[0136] Stagnant core growth indicators: At least one of the following conditions must be met during the above-mentioned period:

[0137] The daily growth rate of pollen coverage was lower than 20% of the expected growth rate of the corresponding stage of the standard development model for five consecutive monitoring periods, and its absolute value was less than 0.5% / day.

[0138] The daily growth rate of fruit powder coverage thickness was lower than 20% of the expected growth rate of the corresponding stage of the standard development model for five consecutive monitoring periods, and its absolute value was less than 0.8 gray units / day.

[0139] Warning issued: "Stunted growth. Suspected nutrient deficiency or root stress. It is recommended to check soil fertility and moisture conditions."

[0140] Rule B, Premature Degradation; This rule is designed to identify damage to the bloom layer caused by acute external injury (such as physical bruising, hail, insect bites) or sudden surface diseases (such as sooty mold).

[0141] Judgment conditions:

[0142] Stage of occurrence: The abnormality occurs before the fruit enters the ripening stage;

[0143] Characteristic mutation: Within two consecutive monitoring periods (e.g., within 48 hours), a drastic and irreversible negative change occurs in the fruit bloom characteristic, satisfying at least one of the following:

[0144] A sharp drop in bloom thickness: The quantitative value of bloom coverage thickness decreased by more than 15% of its previous stable value;

[0145] A sharp increase in bloom uniformity: The quantitative value of bloom uniformity suddenly increases by more than 40%, indicating that the bloom layer is damaged and uneven.

[0146] Irreversible: In the subsequent three monitoring periods, the indicators that plummeted or surged failed to recover to 90% of their pre-mutation values.

[0147] Warning: Premature degradation of bloom. Suspected physical damage or sudden surface disease (such as sooty mold). Immediate field inspection recommended.

[0148] Rule C: Pathological mottling; This rule aims to identify conditions that result in characteristic lesions on the bloom layer caused by locally infectious diseases (such as anthracnose).

[0149] Judgment conditions:

[0150] Macroscopic growth is normal: the time-series curves of bloom coverage and thickness basically conform to the standard development pattern, and the development deviation index may be within the normal range or slightly high.

[0151] Persistently abnormal uniformity: The uniformity index of fruit powder was consistently higher than the 95th percentile of the statistical value of healthy samples at the same developmental stage for five consecutive monitoring periods, indicating that there was a persistent high degree of unevenness in the distribution of fruit powder.

[0152] Lesion feature identification: In the white bloom mask area of ​​blueberry polarized image, multiple low-brightness "dark holes" were continuously detected; specifically quantified as: the total area of ​​pixels with a gray value of less than 50 (range of 0-255) in the mask accounted for more than 5% of the total area of ​​the bloom mask.

[0153] Warning signal: Pathological mottling. This indicates an early risk of suspected fungal infection (such as anthrax). Close observation and consideration of preventative sterilization measures are recommended.

[0154] This invention, through in-depth analysis of abnormal patterns, outputs specific diagnostic conclusions and links them to possible root causes; enabling growers to conduct precise and differentiated field management, greatly enhancing the practical value of the system, and helping users move from identifying problems to solving them, which is a key link in realizing closed-loop management of precision agriculture.

[0155] For fruit bloom time-series characteristics of abnormal development types, the abnormality type is identified and an early warning is issued; for fruit bloom time-series characteristics of normal development types, predictions are made in conjunction with environmental prediction data, the maturity date is determined, and the result is output.

[0156] The process for predicting the maturity date of the fruit powder timing characteristics of the normal development type includes:

[0157] A multivariate regression prediction model is constructed that integrates pollen time-series characteristics and environmental factors. The multivariate regression prediction model uses the pollen time-series characteristics of normal development type as the core input and integrates environmental prediction data. The multivariate regression prediction model is a machine learning model that can handle multiple input variables (independent variables) and predict a continuous output value (dependent variable), such as XGBoost or a neural network with a regression head.

[0158] During the training phase of the multiple regression prediction model, historical maturity data is used to learn the time period required for a normally developing blueberry to evolve to its final mature state, as well as the mapping relationship between the changes in white bloom characteristics and environmental data during the time period.

[0159] Input vector construction: For a fruit determined to be developing normally, the system constructs a long concatenated vector as the model input at the current moment. This vector includes the temporal characteristics of fruit pollen over a past period, as well as key environmental prediction data for the future period, such as average daily temperature, effective accumulated temperature, and total sunshine hours.

[0160] The multivariate regression prediction model is trained offline using historical ripening data. Each sample in the training data contains a concatenated input vector as described above, and its corresponding true label, which is the actual number of days it took for the fruit to ripen from its current state.

[0161] The multivariate regression prediction model consists of four core modules: a two-stream time series encoder, an attention fusion module, a biologically constrained prediction head, and a composite loss function.

[0162] Dual-stream time series encoder: This module is responsible for parallel processing of two different types of time series input data to deeply extract their dynamic features, including biometric encoding stream and environmental factor encoding stream;

[0163] For the biometric encoding stream, the continuously acquired time-tagged white bloom features, namely "bloom temporal features", specifically include time-varying sequence data such as bloom coverage, bloom thickness, and bloom uniformity. A gated recurrent unit (GRU) network is used to capture the inherent long-term evolution patterns and dependencies of these biometric indicators.

[0164] For the environmental factor encoding stream, input future environmental prediction data, such as time series data of average daily temperature, effective accumulated temperature and total sunshine hours for the next few days; use an independent GRU network to extract key external driving patterns from the environmental change sequence.

[0165] Attention Fusion Module: This module is responsible for intelligently fusing information from the two encoders mentioned above; it can determine which future environmental factors (such as light and temperature) contribute the most to maturation prediction at the current stage based on the current biological characteristics (if the fruit development stage), and thus assign them higher weights, achieving intelligent collaborative analysis of internal and external factors.

[0166] Biologically Constrained Prediction Head: This module is the final decision part of the model and is responsible for outputting the prediction results.

[0167] The regression head, composed of a multilayer perceptron, receives a high-dimensional feature vector fused by an attention module and outputs a continuous numerical value representing the predicted time required for blueberries to ripen. Based on this result, it then determines and outputs the specific ripening date.

[0168] Composite loss function: This is a core innovation in the model training process. By introducing prior biological knowledge as a constraint, it ensures the accuracy and reasonableness of the prediction results; it includes mean squared error loss and biological reasonableness loss.

[0169] The mean squared error loss is a standard regression loss that ensures prediction accuracy by minimizing the difference between the predicted number of days and the actual number of mature days in historical data.

[0170] The biological rationality loss refers to the evaluation of the rationality of the future evolution trend of fruit powder characteristics inferred within the model by using a pre-set biological growth model, thereby ensuring the biological rationality of the prediction process.

[0171] The biological growth model can employ a logistic function, the parameters of which are obtained by nonlinearly fitting a large amount of historical time-series data of normally developing blueberries; the biological rationality loss is the root mean square error between the model's predicted future feature evolution trajectory and the fitted curve.

[0172] In the prediction phase of the multiple regression prediction model, the current fruit pollen time-series characteristics are received, and combined with the acquired environmental prediction data, the future evolution data of the fruit pollen time-series characteristics are dynamically deduced; the time required for the future evolution data to reach the preset maturity threshold is calculated, and the maturity date is determined.

[0173] This invention constructs a multivariate regression prediction model that integrates fruit powder temporal characteristics and environmental factors. By combining an internal biological model with an external environmental model, it accurately determines the ripening date of blueberries and provides a dynamically adaptive harvesting period prediction. This can effectively reduce harvesting risks, minimize losses, and maximize market value.

[0174] Furthermore, it also includes steps to provide decision support for robotic harvesting:

[0175] For fruit powder time-series characteristics that are determined to be of normal development type and have reached maturity date, the system activates a clamping force prediction process.

[0176] The process first extracts a set of multimodal visual features related to the physical hardness of the fruit. In addition to the fruit bloom coverage, fruit bloom thickness, fruit bloom uniformity and fruit peel background color quantification values, these features also include the size and outline regularity obtained by analyzing the fruit outline.

[0177] Subsequently, the multimodal visual features are input into a regression prediction model that has been pre-trained using "visual features - true hardness" paired data;

[0178] Finally, the model outputs an optimal gripping force parameter and transmits this parameter to the control unit of the robot harvesting system to guide it in performing a low-damage grasping action on the blueberry fruit.

[0179] Example 2:

[0180] This invention also proposes an image recognition-based blueberry fruit detection system to implement the method described in Embodiment 1. The framework structure of each module of the system is as follows: Figure 3 As shown, it includes: an image acquisition module, an Apple fan recognition module, a feature quantization module, a type classification module, and a judgment output module.

[0181] The image acquisition module uses cross-polarization imaging technology to monitor blueberry fruits and obtain blueberry polarization images and standard blueberry images.

[0182] Preferably, in the process of imaging and monitoring blueberry fruits using cross-polarization imaging technology, the monitoring equipment includes an intelligent monitoring arm deployed on an automatic inspection track, which integrates a 1920x1080 resolution industrial camera, a 50mm fixed-focus lens, an electric polarizer group, and a 6500K ring LED active light source.

[0183] Move the monitoring equipment to the blueberry cluster and perform the imaging task: acquire a standard blueberry image, control the analyzer to be parallel with the polarizer, flash the LED, and capture a 1920x1080 pixel RGB image; acquire a blueberry polarized image, rotate the analyzer 90° to the orthogonal position, flash the LED, and capture a cross-polarized image.

[0184] The fruit bloom recognition module constructs a blueberry stage recognition model to recognize the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, it selects an image segmentation strategy to adaptively segment the blueberry polarization image and recognizes the white bloom mask.

[0185] Developmental stage determination: The blueberry fruit area in the standard blueberry image is fed into the pre-trained blueberry stage recognition model. The model analyzes the outline, size and apparent color of the blueberry fruit in the image, outputs a probability vector, and determines its developmental stage as the color-changing stage.

[0186] For example, the blueberry stage recognition model consists of 3 convolutional blocks and 2 fully connected layers; each convolutional block contains a 3x3 convolutional layer (ReLU activation function) and a 2x2 max pooling layer; the last fully connected layer is followed by a Softmax activation function to output the probabilities of the 3 classes. During training, the Adam optimizer is used, with an initial learning rate set to 0.001, a batch size of 32, and a total of 100 training epochs.

[0187] Based on the developmental stage determination during the color-changing period, a local adaptive thresholding algorithm was selected to process the blueberry polarized image. This algorithm dynamically calculates the local threshold based on the global distribution within the neighborhood of each pixel, with the neighborhood size set to 15x15 pixels. This algorithm successfully and accurately segmented the early bloom areas with uneven brightness and mottled shapes from the darker fruit skin background, generating a bloom mask. The grayscale image of a mature blueberry is shown below. Figure 4 As shown.

[0188] For example, for blueberry polarization images during the color change period, a Gaussian weighted sum adaptive thresholding algorithm is used as a local adaptive thresholding algorithm for segmentation; the neighborhood size of the algorithm is set to 15x15 pixels, and the constant C value is 2 subtracted from the calculated Gaussian weighted sum.

[0189] The feature quantization module constructs a fruit bloom feature quantification model to identify blueberry polarized images, standard blueberry images, and white bloom masks, and obtain white bloom features, specifically including fruit bloom coverage, fruit bloom coverage thickness, fruit bloom uniformity, and fruit skin background color quantification values.

[0190] A white bloom mask is applied to the total pixel area of ​​the fruit, and the ratio of the number of pixels within the mask to the total number of pixels on the fruit is calculated to obtain the bloom coverage rate. The pixel grayscale values ​​corresponding to the white bloom mask area are extracted from the blueberry polarized image, and the average value of the pixel grayscale values ​​is calculated as a quantitative indicator of bloom coverage thickness. The standard deviation of the pixel grayscale values ​​within the white bloom mask area is calculated in the blueberry polarized image to assess the dispersion of bloom distribution, serving as the bloom uniformity. The pixel area corresponding to the white bloom mask is located in the blueberry standard image, and the exposed peel area is obtained by removing pixels. The chromaticity values ​​of the exposed peel area in the color space are extracted and calculated, and quantized as the peel background color quantization value. Based on the bloom coverage rate, bloom coverage thickness, bloom uniformity, and peel background color quantization value, the white bloom characteristics are obtained.

[0191] The type classification module continuously acquires the white bloom fruit powder features with time tags to obtain the fruit powder temporal features; constructs a blueberry dynamic development model to identify the fruit powder temporal features, analyzes the synergistic relationship between the evolution of fruit powder features and the background color of the fruit peel, and compares and analyzes them with standard temporal features to classify them into normal development type and abnormal development type.

[0192] The blueberry bloom feature is continuously acquired with time tags to obtain bloom time-series features. The dynamic evolution trend of these bloom time-series features is analyzed using a pre-trained blueberry dynamic development model. Similarity matching and deviation calculation are performed between this model and the learned standard development pattern to obtain a development deviation index. Then, based on a preset deviation threshold, normal and abnormal development types are determined. The blueberry dynamic development model uses a two-layer GRU network with 128 hidden units, followed by a fully connected layer for output. During training, the learning rate is set to 0.005.

[0193] The output module determines the abnormality type and outputs a warning for fruit pollen time-series characteristics of abnormal development types; for fruit pollen time-series characteristics of normal development types, it combines environmental prediction data to predict the maturity date and output it.

[0194] Once the fruit pollen time sequence characteristics are determined to be an abnormal development type, a diagnostic module based on rule and pattern recognition will be activated; the diagnostic module analyzes the deviation patterns of the fruit pollen time sequence characteristics from the standard development pattern, and identifies the abnormal type based on the deviation patterns.

[0195] Once the fruit pollen time-series characteristics are determined to be of normal development type, the current fruit pollen time-series characteristics are received, and combined with the acquired environmental prediction data, the future evolution data of the fruit pollen time-series characteristics are dynamically deduced; the time required for the future evolution data to reach the preset maturity threshold is calculated, and the maturity date is determined.

[0196] For example, the maturity threshold can be set to simultaneously satisfy the following conditions:

[0197] The pollen coverage reaches or exceeds a stable threshold, such as 95%;

[0198] The powder coating thickness reaches or exceeds a stable threshold, such as 210 (in an 8-bit grayscale image of 0-255);

[0199] The quantified value of the peel background color enters and stabilizes within a specific range that represents the fully ripe color of a particular variety, such as dark blue or bluish-purple. For example, the H value is in the range of [230, 255] (based on a hue wheel of 0-360 degrees mapped to an 8-bit integer).

[0200] The development rate tends to plateau: the daily growth rate of bloom coverage and thickness (the slope of the time series curve) is close to zero for three consecutive monitoring periods (e.g., the absolute value of the daily growth rate is less than 1%).

[0201] Information warnings are issued based on the abnormality type of abnormal development and the maturity date of normal development.

[0202] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A blueberry fruit detection method based on image recognition, characterized by, include: Blueberry fruits were imaged and monitored using cross-polarization imaging technology to obtain polarized images and standard images of blueberries. A blueberry development stage recognition model is constructed to identify the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, an image segmentation strategy is selected to adaptively segment the blueberry polarization image and identify the white bloom mask; A fruit bloom feature quantification model was constructed to identify blueberry polarization images, standard blueberry images, and white bloom masks to obtain white bloom features, specifically including fruit bloom coverage, fruit bloom thickness, fruit bloom uniformity, and fruit skin background color quantification values. Continuously acquire time-tagged white bloom features to obtain bloom time-series features; construct a blueberry dynamic development model to identify the bloom time-series features, analyze the synergistic relationship between the evolution of bloom features and the background color of the peel, and compare and analyze with standard time-series features to classify them into normal development type and abnormal development type. For the time-series characteristics of fruit powder with abnormal development, determine the abnormality type and output an early warning; For the time-series characteristics of fruit powder in normal development types, predictions are made in conjunction with environmental prediction data to determine the maturity date and output it.

2. The blueberry fruit detection method based on image recognition according to claim 1, characterized in that, The steps for imaging and monitoring blueberry fruit based on cross-polarization imaging technology include: An active light source system is used, which integrates a first linear polarizer to emit polarized light onto the blueberry fruit; at the imaging end, a high-resolution image sensor is used, and a second linear polarizer is deployed in front of the lens as an analyzer. The polarization direction of the analyzer is controlled to be parallel to the polarization direction of the first linear polarizer, and a standard blueberry image containing specular reflection light from the fruit surface is acquired to identify the outline, size and apparent color of the blueberry fruit. The polarization direction of the analyzer is rotated by 90 degrees so that it is orthogonal to the polarization direction of the first linear polarizer. Under cross-polarization, light that has been depolarized by multiple scatterings within the white bloom is collected to obtain a blueberry polarization image that highlights the spatial distribution and density of the white bloom.

3. The blueberry fruit detection method based on image recognition according to claim 1, characterized in that, The steps to build a blueberry stage identification model and determine the blueberry development stage include: Standard images covering all stages of blueberries from immature to fully ripe were collected and labeled according to the outline, size and apparent color of the blueberry fruit to create a training dataset with three label types: young fruit stage, color change stage and fully ripe stage. A deep learning-based convolutional neural network architecture is used as the foundation of the blueberry stage recognition model; supervised learning training is performed on the training dataset to enable the model to learn to extract multi-dimensional features of color, texture and shape from standard images; During the identification phase, the collected standard blueberry images are input into the trained blueberry stage identification model. Through forward propagation calculation, the probability of each developmental stage is output, and the stage with the highest probability value is determined as the blueberry developmental stage.

4. The blueberry fruit detection method based on image recognition according to claim 3, characterized in that, The process of adaptive segmentation of blueberry polarization images based on image segmentation strategies selected according to blueberry development stages includes: For blueberry polarized images in the young fruit stage of blueberry development, image segmentation is performed based on a preset severity threshold; For blueberry polarized images in the color-changing stage of blueberry development, a local adaptive thresholding algorithm is used for image segmentation. The local adaptive thresholding algorithm dynamically calculates the local threshold based on the global distribution within the neighborhood of each pixel. For polarized images of blueberries at the fully ripe stage of development, the Otsu adaptive thresholding algorithm was used for image segmentation.

5. The blueberry fruit detection method based on image recognition according to claim 1, characterized in that, The characteristics of the white bloom fruit powder include bloom coverage, bloom thickness, bloom uniformity, and quantitative value of the fruit peel background color; The fruit bloom coverage rate is obtained by applying a white bloom mask to the total pixel area of ​​the fruit and calculating the ratio of the number of pixels in the mask to the total number of pixels in the fruit. The fruit bloom coverage thickness is determined by extracting the pixel grayscale values ​​of the corresponding region of the white bloom mask in the blueberry polarization image, and calculating the average value of the pixel grayscale values ​​as a quantitative indicator of the fruit bloom coverage thickness. The uniformity of the fruit bloom is determined by calculating the standard deviation of pixel gray values ​​within the white bloom mask area in a blueberry polarized image to assess the dispersion of the fruit bloom distribution, which is then used as the uniformity of the fruit bloom. The quantization value of the fruit peel background color is obtained by locating the pixel area corresponding to the white bloom mask in the blueberry standard image and removing it to obtain the exposed fruit peel area; the chromaticity value of the exposed fruit peel area in the color space is extracted and calculated, and quantized into the fruit peel background color quantization value.

6. The blueberry fruit detection method based on image recognition according to claim 1, characterized in that, The steps for constructing a dynamic developmental model of blueberries and classifying developmental types include: A recurrent neural network capable of processing time-series data was adopted as the core architecture of the blueberry dynamic development model. Using historical monitoring data of blueberry bloom time-series feature samples marked as normal development, the blueberry dynamic development model was trained to learn and encode the inherent laws and standard development patterns of bloom coverage, bloom thickness, bloom uniformity and peel background color quantification values ​​co-evolving over time. During model recognition, the dynamic evolution trend of fruit powder time sequence characteristics is analyzed through the blueberry dynamic development model, and similarity matching and deviation calculation are performed with the internally learned standard development pattern to obtain the development deviation index; then, based on the preset deviation threshold, the normal development type and abnormal development type are determined.

7. The image recognition-based blueberry fruit detection method according to claim 6, characterized in that: Once the fruit powder time sequence characteristics are determined to be an abnormal development type, the rule-based and pattern recognition-based diagnostic module will be activated. The diagnostic module analyzes deviation patterns in fruit pollen time-series characteristics that deviate from the standard developmental pattern, and identifies abnormality types based on these deviation patterns.

8. The blueberry fruit detection method based on image recognition according to claim 1, characterized in that, The process for predicting the maturity date of the fruit powder timing characteristics of the normal development type includes: A multivariate regression prediction model integrating pollen temporal characteristics and environmental factors was constructed. The multivariate regression prediction model uses the pollen temporal characteristics of normal development type as the core input and integrates environmental prediction data. During the training phase of the multiple regression prediction model, historical maturity data is used to learn the time period required for a normally developing blueberry to evolve to its final mature state, as well as the mapping relationship between the changes in white bloom characteristics and environmental data during the time period. In the prediction phase of the multiple regression prediction model, the current fruit pollen time-series characteristics are received, and combined with the acquired environmental prediction data, the future evolution data of the fruit pollen time-series characteristics are dynamically deduced; the time required for the future evolution data to reach the preset maturity threshold is calculated, and the maturity date is determined.

9. A blueberry fruit detection system based on image recognition, for implementing the blueberry fruit detection method based on image recognition of claim 1, characterized in that, include: The image acquisition module uses cross-polarization imaging technology to monitor blueberry fruits and obtain blueberry polarization images and standard blueberry images. The fruit bloom recognition module constructs a blueberry stage recognition model to recognize the standard blueberry image and determine the blueberry development stage; based on the blueberry development stage, it selects an image segmentation strategy to adaptively segment the blueberry polarization image and recognizes the white bloom mask. The feature quantization module constructs a fruit bloom feature quantification model to identify blueberry polarized images, standard blueberry images, and white bloom masks, and obtain white bloom features, specifically including fruit bloom coverage, fruit bloom coverage thickness, fruit bloom uniformity, and fruit skin background color quantification values. The type classification module continuously acquires the white bloom fruit powder features with time tags to obtain the fruit powder temporal features; constructs a blueberry dynamic development model to identify the fruit powder temporal features, analyzes the synergistic relationship between the evolution of fruit powder features and the background color of the fruit peel, and compares and analyzes them with standard temporal features to classify them into normal development type and abnormal development type. The output module determines the abnormality type and outputs a warning for the time-series characteristics of fruit pollen with abnormal development. For the time-series characteristics of fruit powder in normal development types, predictions are made in conjunction with environmental prediction data to determine the maturity date and output it.

Images