Method and system for navel orange defect detection

Abstract

The application provides a navel orange defect detection method and system, and belongs to the technical field of navel orange detection. The method comprises the following steps: acquiring a multi-scale visual feature map; extracting multi-dimensional physiological parameters of the navel orange and performing quantitative stratification; sequentially performing layer-by-layer reinforcement on the visual feature map by using shallow texture mechanism features, middle layer shape mechanism features, high layer spatial distribution mechanism features and top layer hidden mold mechanism features to generate layer-by-layer reinforcement feature maps; performing adaptive mechanism global fusion on the layer-by-layer reinforcement feature maps and the visual feature map to generate a mechanism-enhanced hybrid feature map; performing mechanism constraint on the mechanism-enhanced hybrid feature map in the space and shape of the navel orange by using an adaptive suppression constraint term to obtain a final feature map; and performing defect detection by using a shared convolution strategy to obtain a defect detection result. The application has the advantages of mechanism support, strong defect recognition pertinence and strong anti-interference capability.

Description

Technical Field

[0001] This invention relates to the field of navel orange testing technology, and more specifically to a method and system for detecting defects in navel oranges. Background Technology

[0002] Navel oranges are a distinctive and advantageous economic fruit crop in southern my country, with planting area and annual output ranking among the world's top. Post-harvest appearance quality is a core indicator determining its commercial grade, market price, and storage period. Throughout the entire process of growth, harvesting, transportation, and storage, navel oranges are susceptible to multiple factors, including pests and diseases, extreme field environments, human and mechanical collisions, and the growth of storage molds. These factors can lead to various surface defects such as canker, mechanical damage, sunburn, and mold, severely reducing the commercial value of the fruit and even causing mass fruit rot, resulting in huge economic losses. Therefore, achieving rapid and accurate detection of navel orange defects is a key technological support for the intelligent grading, quality improvement, and efficiency enhancement of the navel orange industry.

[0003] Currently, navel orange defect detection is mainly divided into three types: manual sorting, traditional machine vision inspection, and deep learning intelligent inspection. The existing technical problems are prominent: manual sorting relies on human observation, which has problems such as low efficiency, high labor costs, large subjective errors, and long-term fatigue leading to missed or false detections, and cannot meet the needs of large-scale industrial production; traditional machine vision inspection relies on texture and grayscale feature extraction with fixed thresholds, which can only identify large-area defects with significant features, and has extremely poor adaptability to early and weak defects with irregular shapes and blurred edges. Moreover, changes in ambient lighting and background clutter can cause the algorithm to completely fail, resulting in extremely low environmental robustness.

[0004] With the development of machine vision and deep learning technologies, target detection algorithms, represented by the YOLO series, have been gradually applied to fruit defect detection and are now widely used in the field. Current YOLO-based target detection algorithms rely on a simple combination and optimization of a general lightweight backbone, attention modules, and feature fusion structures. They only fit defect features at the data level, completely detaching from the specific agricultural mechanisms of navel orange fruit's physiological structure, defect pathological evolution, and field damage patterns. They fail to explore key prior information such as the microscopic texture of the peel's oil glands, the specificity of defect spatial distribution, the morphological evolution differences of different diseases, and the essential mechanism differences between inherent organs and defects.

[0005] This results in the YOLO model having a strong tendency to blindly extract features, making it unable to accurately distinguish between subtle defects, irregular defects, and interfering features such as fruit stems, fruit navels, and natural textures. It also leads to serious missed detections and false detections of low-discrimination defects such as early latent mold, minor mechanical damage, and shallow ulcers. At the same time, the purely data-driven model has extremely poor interpretability, and feature learning lacks clear physical and agricultural basis. Model optimization relies on trial and error, has weak generalization ability, and is difficult to adapt to complex working conditions such as field light fluctuations, fruit posture changes, and cross-batch fruit peel color differences. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting defects in navel oranges, so as to at least solve the problems of lack of mechanistic support, weak defect identification specificity, and poor anti-interference ability of the YOLO model when performing defect detection.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for detecting defects in navel oranges, the method comprising:

[0008] After performing multi-scale feature extraction on the navel orange image using the YOLO11 model, the multi-scale visual feature map output by the neck network of the YOLO11 model is obtained.

[0009] Multidimensional physiological parameters of navel oranges were extracted and quantitative stratification was performed to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features.

[0010] By utilizing the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features, the visual feature map is enhanced layer by layer to generate enhanced feature maps at each level.

[0011] Adaptive mechanism-based global fusion is performed between the enhanced feature maps at each level and the visual feature maps to generate a hybrid feature map with enhanced mechanism.

[0012] An adaptive suppression constraint term for the navel orange interference target is constructed. The adaptive suppression constraint term is used to perform mechanistic constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map.

[0013] The final feature map is input into the head network of the YOLO11 model, and defect detection is performed using a shared convolution strategy to obtain the defect detection results.

[0014] Preferably, the YOLO11 model includes a backbone network, a neck network, and a head network;

[0015] The backbone network adopts the StarNet structure to perform multi-scale feature extraction on the navel orange image, resulting in four initial feature maps at different scales.

[0016] The neck network adopts a BiFPN structure to perform different scale fusion and nonlinear enhancement on the initial feature maps of four different scales to obtain fused features of three scales, and uses the fused features of three scales as the visual feature map.

[0017] The head network is used to perform defect detection on the final feature map using a shared convolution strategy to obtain defect detection results.

[0018] Preferably, the ends of the backbone network are embedded with parameterless attention units, which are used to perform attention enhancement processing on the initial feature map at the smallest scale.

[0019] Preferably, the process involves extracting multidimensional physiological parameters of navel oranges and performing quantitative stratification, including:

[0020] The effective number of oil glands, total number of oil glands, perimeter and enclosed area of ​​defects, geometric center coordinates of the fruit and equivalent radius of the fruit, and the mean fluorescence gray value of the background area, healthy peel area and suspected moldy area in the fluorescence image were collected.

[0021] Based on the number of effective oil glands and the total number of oil glands, the oil gland integrity index is calculated, the gray-scale dispersion of local peel areas is statistically analyzed, and the texture disorder coefficient is calculated based on the oil gland integrity index and the gray-scale dispersion. The shallow texture mechanism characteristics are constructed using the oil gland integrity index and the texture disorder coefficient.

[0022] Based on the perimeter of the defect's outline and the area it encloses, the defect irregularity is calculated. A normalization mapping is then performed on the defect irregularity to obtain the morphological regularity coefficient. The mid-layer morphological mechanism features are then constructed using the defect irregularity and the morphological regularity coefficient.

[0023] Based on the geometric center coordinates of the fruit and the equivalent radius of the fruit, the normalized distance of the region is calculated, and the defect location factor is solved based on the preset region weight factor. The region fitness factor is solved according to the mechanism threshold discrimination rule. The high-level spatial distribution mechanism features are constructed using the defect location factor and the region fitness factor.

[0024] The fluorescence intensity index is calculated based on the average fluorescence grayscale values ​​of the background area, healthy peel area, and suspected moldy area in the fluorescence image; the area of ​​abnormal fluorescence response area and the total effective area of ​​the fruit are statistically analyzed to determine the proportion of infected area; the mold infection degree coefficient is determined based on the fluorescence intensity index and the proportion of infected area; and the top-level latent mold mechanism characteristics are constructed using the fluorescence intensity index and the mold infection degree coefficient.

[0025] Preferably, the enhanced feature maps at each level include: microscopic feature maps, morphological optimization feature maps, effective feature maps, and cross-modal feature maps; the multi-scale visual feature maps are sequentially enhanced using shallow texture mechanism features, mid-level morphological mechanism features, high-level spatial distribution mechanism features, and top-level latent mold mechanism features, including:

[0026] Based on the shallow texture mechanism features, micro-texture enhancement is performed on the visual feature map to generate a micro-feature map;

[0027] Based on the mid-level morphological mechanism characteristics, differential morphological enhancement is performed on the micro-feature map to generate a morphologically optimized feature map;

[0028] Based on the spatial distribution mechanism characteristics of high-level structures, pseudo-defect filtering is performed on the morphological optimization feature map to generate an effective feature map.

[0029] Based on the top-level latent mold growth mechanism characteristics, latent defect enhancement is performed on the effective feature map to generate cross-modal feature maps.

[0030] Preferably, the adaptive inhibition constraint term includes a morphological mechanism constraint term created by the morphological regularity coefficient and a spatial distribution mechanism constraint term created by the region fit factor; the morphological mechanism constraint term is used to perform inhibition on the region where the morphologically regular interference target is located, and the spatial distribution mechanism constraint term is used to perform inhibition on the region that conforms to the inherent organ distribution law.

[0031] Preferably, the mechanism constraints on the execution morphology and space of the hybrid feature map based on constraint terms include:

[0032] Extract the pixel-level binary mask of the interfering target;

[0033] Based on pixel-level binary masks, spatial distribution mechanism constraints, and morphological mechanism constraints, the computer adaptive suppression weight coefficients are calculated.

[0034] Based on the mechanism-adaptive suppression weight coefficient, the mechanism-enhanced hybrid feature map is subjected to mechanism-flexible feature purification to generate a purified defect feature map, which is then used as the final feature map.

[0035] Secondly, the present invention provides a navel orange defect detection system for implementing the above-mentioned navel orange defect detection method, the system comprising:

[0036] The feature acquisition module is used to obtain the multi-scale visual feature map output by the neck network of the YOLO11 model after the YOLO11 model performs multi-scale feature extraction on the navel orange image.

[0037] The mechanism quantification module is used to extract multi-dimensional physiological parameters of navel oranges and perform quantification stratification to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features.

[0038] The feature enhancement module is used to sequentially enhance the visual feature map by utilizing the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features, generating enhanced feature maps at each level.

[0039] The feature fusion module is used to perform adaptive global fusion of enhanced feature maps at each level with visual feature maps to generate a hybrid feature map with enhanced mechanism.

[0040] The mechanism constraint module is used to construct an adaptive suppression constraint term for the navel orange interference target. The adaptive suppression constraint term is used to perform mechanism constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map.

[0041] The defect detection module is used to input the final feature map into the head network of the YOLO11 model, perform defect detection using a shared convolution strategy, and obtain the defect detection results.

[0042] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the navel orange defect detection method in Embodiment 1.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention quantifies the multi-dimensional physiological parameters of the navel orange's physiological structure to distinguish the differentiated characteristics of different defects and different degrees of disease, which can provide prior support for the accurate identification of navel orange defects and improve the accuracy of defect identification.

[0045] 2. This invention adopts a hierarchical enhancement strategy to adaptively match the feature requirements of defects at different scales and developmental stages, thus solving the problem of insufficient feature extraction and serious feature confusion of the YOLO11 model for weak defects, small-sized defects and irregular defects.

[0046] 3. This invention builds an adaptive inhibition constraint term by relying on the differences in physiological, distribution, and morphological mechanisms between defects and inherent organs. This term applies morphological and spatial mechanism constraints to the hybrid feature map with enhanced mechanism, which can accurately eliminate false detection interference from interfering targets such as fruit stalks, fruit navels, and natural textures.

[0047] 4. The head network of this invention utilizes a shared convolution strategy, which can improve the utilization rate of semantic information in the network and effectively balance the detection accuracy of the YOLO11 model with the requirements of lightweight deployment. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart of the navel orange defect detection method provided by the present invention;

[0050] Figure 2 This is a schematic diagram of the architecture of the head network of the YOLO11 model of the present invention;

[0051] Figure 3This is a block diagram of the navel orange defect detection system provided by the present invention. Detailed Implementation

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0053] Example 1

[0054] This embodiment provides a method for detecting defects in navel oranges, such as... Figure 1 As shown, the method includes the following steps:

[0055] Step S100: After the YOLO11 model performs multi-scale feature extraction on the navel orange image, obtain the multi-scale visual feature map output by the neck network of the YOLO11 model.

[0056] In this embodiment, the YOLO11 model includes a backbone network, a neck network, and a head network, which are sequentially connected. To address the issues of lack of mechanistic support, weak defect identification specificity, and poor anti-interference ability in defect detection by the YOLO model, this embodiment extracts the multi-scale visual feature map output by the neck network separately. After processing in steps S200 to S500, the final feature map is obtained, and then input into the neck network to perform defect detection, thereby overcoming the aforementioned problems.

[0057] As a further optimization of this embodiment, in order to enable the backbone network to complete the preliminary extraction of basic visual features such as navel orange peel texture and defect edges under the premise of extremely low parameters and extremely low computational cost, this embodiment improves the architecture of the YOLO11 model backbone network, that is, adopts the StarNet structure for the backbone network.

[0058] The StarNet architecture contains four Star Blocks modules. Each Star Block uses star operations, which fuse features from different subspaces through element-wise multiplication. Each Star Block outputs an initial feature map at one scale, resulting in four initial feature maps. This achieves implicit high-dimensional feature mapping while maintaining lightweight design and basic feature extraction capabilities. It can perform preliminary extraction of basic visual features such as navel orange peel texture and defect edges with extremely low parameters and computational cost.

[0059] In this embodiment, a depthwise convolution (DW-Conv) is added to the end of each Star Blocks module, and batch normalization (BN) is placed after the depthwise convolution to enhance feature processing efficiency. In these convolutional operations, the dilation factor is set to 4, thereby expanding the receptive field without excessively increasing computational burden. Without changing the convolutional layers, the Star Blocks module replaces the four C3K2 modules in the original backbone network of the YOLO11 model, maintaining detection accuracy as much as possible while avoiding complex network architecture design, allowing the model to efficiently acquire richer feature information.

[0060] The formula for calculating star operations is as follows:

[0061] (1)

[0062] In equation (1), Input feature maps to the backbone network; The weight matrix is ​​a learnable linear transformation. For bias terms; This is an element-wise multiplication operation. This structure can efficiently extract the texture and defect edge details of navel orange peel while significantly compressing model parameters.

[0063] As a further optimization of this embodiment, a parameterless attention unit is embedded at the end of the backbone network. The parameterless attention unit is used to perform attention enhancement processing on the initial feature map of the smallest scale.

[0064] Specifically, a parameterless attention unit is connected after the C2PSA module in the backbone network. The parameterless attention unit performs self-attention and 3D attention weighting on the feature map output by the C2PSA module (which is the initial feature map at the smallest scale) to make the semantic information stronger and more discriminative. The parameterless attention unit is a separate output, which is equivalent to providing a strengthened high-level semantic feature for the neck network part. It helps to distinguish defects from easily confused areas such as fruit stems and fruit navels, enabling the model to pay attention to targets with four different receptive fields at the same time, thereby improving the recall rate and localization accuracy of small targets.

[0065] After the backbone network outputs four initial feature maps, they are input into the neck network. The neck network performs different scale fusion and nonlinear enhancement on the four initial feature maps at different scales to obtain fused features at three scales. The fused features at three scales are used as visual feature maps.

[0066] In this embodiment, the original neck network of the YOLO11 model uses a PAN (Feature Pyramid Network) structure combined with a C3K2 module for multi-scale feature fusion. This structure uses a simple addition or splicing method for feature fusion, and assumes that features at different levels contribute equally to the fusion result. It cannot dynamically adjust the fusion weights according to the scale and shape of the defect target.

[0067] To address the aforementioned issues, the neck network in this embodiment employs a BiFPN (Bi-FPN Multi-Scale Feature Fusion Network) structure. This structure constructs bidirectional paths that run in parallel from top to bottom and bottom to top, allowing features from each layer to be repeatedly transferred and fused across multiple scales. It introduces learnable weight parameters (using fast normalized fusion), enabling the network to automatically learn the importance of feature maps at different scales based on training data. For navel orange defect detection, this structure allows high-resolution feature maps containing small defects to receive higher fusion weights, thereby improving recall. The neck network in this embodiment outputs visual feature maps at three different scales.

[0068] Simultaneously, Star Blocks modules are embedded again within each C3K2 module to map the input to an extremely high-dimensional implicit feature space through star operations. The equivalent dimension increases exponentially with the number of layers, achieving non-linear enhancement. Embedding Star Blocks modules into the BiFPN fusion nodes is equivalent to introducing a powerful non-linear transformer at each fusion step, enabling the network to capture extremely subtle grayscale or texture differences between the defect region and the background. Therefore, the improved neck network can compensate for the model's shortcomings in handling high-dimensional non-linear features and enhance the model's feature fusion capabilities.

[0069] In this embodiment, a CMOS industrial camera is used to acquire images of navel oranges. The image resolution is 1920×1080, and the storage format is JPG to obtain an RGB image. The RGB image is used as the navel orange image to be detected.

[0070] Step S200: Extract multi-dimensional physiological parameters of navel oranges and perform quantitative stratification to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features.

[0071] In this embodiment, the formation of various defects in navel oranges is the result of the combined effects of physiological damage, pathogen infection, and environmental stress. The causes of different defects determine their unique microscopic texture, macroscopic morphology, spatial distribution, and evolutionary characteristics. Moreover, the same defect exhibits significant characteristic differences in its early, middle, and late stages, demonstrating strong mechanistic specificity. The fruit stalk and navel, as inherent physiological organs of the navel orange, possess stable physiological structures, morphological characteristics, and distribution patterns, which are fundamentally different from disease and damage defects. This is a core agricultural priori for distinguishing defects from disturbances. This embodiment systematically reviews the differentiated mechanistic rules of four types of common defects from five dimensions: fruit physiological basis, defect causes, developmental stages, appearance characteristics, and spatial distribution. It refines the characteristic differences of different disease severity levels and improves the mechanistic discrimination boundary between defects and inherent organs. The specific mechanistic comparison rules are shown in Table 1.

[0072] Table 1. Comparison of Typical Defect Mechanisms in Navel Oranges

[0073] Mechanical injury Physical collisions and squeezing during harvesting, transportation, and sorting can cause damage to the fruit peel. It has no fixed outline, is extremely irregular in shape, and has blurred edges with no obvious boundaries. The oil glands in the peel rupture, the waxy layer peels off, and the surface has uneven gray distribution and no fixed texture pattern. It is mostly concentrated on the equatorial side of navel oranges, and occurs very rarely in the stem and navel areas. peptic ulcer Xanthomonas citrus infects fruit peel cells, causing local tissue proliferation and necrosis. The lesion is nearly round or oval, with a central depression and raised edges, and clear and regular borders. The lesion surface is rough, with ring-shaped textures, and is dark brown in color; the texture features are fixed. The disease is concentrated on the sun-facing side of the fruit and around the stem end, with very little disease on the shaded side. Mold (blue / green mold) Post-harvest storage in a humid environment allows Penicillium and Pseudomonas molds to invade and multiply through damage to the fruit peel. Initially, the lesions appear as irregular, dark, water-soaked spots. Later, the fungal hyphae spread, and the central depression of the lesion expands. Early stages show faint textures and gradual grayscale changes, with significant fluorescence response under ultraviolet light; later stages show hyphal textures. It is most common in the sunken area of ​​the fruit navel and in areas where there is existing mechanical damage. Sunburn defects Direct sunlight and high temperatures in the field cause necrosis and water loss in the fruit peel epidermal cells. Large-area, patchy distribution with blurred boundaries and no fixed geometric shape. The peel has a hardened and rough texture, with a large amount of pigment deposited and the oil glands completely necrotic and disappeared. It is only distributed on the unilateral surface of the fruit that is exposed to sunlight for a long time.

[0074] Table 1 shows that the degree of texture damage, morphological regularity, distribution area, and evolution trend of various defects all follow fixed agricultural pathological patterns and exhibit clear stage differences. In contrast, the fruit stalk and blossom end are normal physiological structures of the fruit, without defects such as oil gland damage, tissue necrosis, or mycelial infection, and their morphology and distribution are stable. The YOLO11 model does not utilize this hierarchical mechanism difference and relies solely on visual pixel features for discrimination, making it prone to misjudgments of "visually similar but mechanistically inconsistent" defects. It cannot distinguish between early, subtle defects and normal fruit peel texture, or between minor diseases and naturally sunken blossom ends, thus limiting detection accuracy and stability.

[0075] In this embodiment, based on the typical defect mechanism characteristics in Table 1, multi-dimensional physiological parameters of navel oranges are collected. By quantifying these multi-dimensional physiological parameters, the differentiated characteristics of different defects and different degrees of disease can be distinguished, providing prior support for the accurate identification of navel orange defects and improving the accuracy of defect identification.

[0076] Specifically, the multi-dimensional physiological parameters include: the number of effective oil glands, the total number of oil glands, the perimeter and enclosed area of ​​the defect, the geometric center coordinates of the navel orange fruit and the equivalent radius of the fruit, the mean fluorescence gray value of the background area, healthy peel area and suspected moldy area in the fluorescence image of the navel orange.

[0077] Therefore, multidimensional physiological parameters of navel oranges were collected and quantitative stratification was performed, including:

[0078] Step 1: Collect the effective number of oil glands, total number of oil glands, perimeter and enclosed area of ​​defects, geometric center coordinates of the fruit and equivalent radius of the fruit, and the mean fluorescence gray value of the background area, healthy peel area and suspected moldy area in the fluorescence image of the navel orange.

[0079] Step 2: Calculate the oil gland integrity index based on the number of effective oil glands and the total number of oil glands, and statistically analyze the gray-scale dispersion of local peel areas. Calculate the texture disorder coefficient based on the oil gland integrity index and the gray-scale dispersion. Construct shallow texture mechanism characteristics using the oil gland integrity index and the texture disorder coefficient. The oil gland integrity index is used to quantify the degree of peel texture damage, and the texture disorder coefficient is used to characterize regional texture disorder and the degree of oil gland damage.

[0080] The specific steps are as follows:

[0081] Step 2.1: Perform foreground segmentation on the navel orange image and remove the background area; convert it to a grayscale image and use Gaussian blur to remove subtle noise from the peel; extract the dark dot-shaped oil gland targets on the peel surface through adaptive local threshold segmentation and morphological opening operation, filter the oil gland area and roundness constraints, remove dust, scratches and noise interference, and complete the accurate differentiation and pixel statistics of oil gland targets.

[0082] Step 2.2: Within the effective area of ​​the fruit peel, count the number of healthy oil glands that are intact, have clear boundaries, and are nearly circular in shape. Based on the effective area of ​​the fruit peel and the distribution density of oil glands in a standard healthy fruit peel, the theoretical total number of oil glands in the region is calculated. .

[0083] Step 2.3: Extract the coordinates of all valid oil gland centers, calculate the variance of the coordinate distribution, and characterize the uniformity of oil gland distribution.

[0084] (2)

[0085] In equation (2), The variance of the coordinate distribution. The coordinates of the center of a single sebaceous gland, The coordinates of the mean center of the sebaceous glands in the region are: The total number of effective oil glands.

[0086] Step 2.4: Calculate the Oil Gland Integrity Index (OGI) based on the number of healthy oil glands, the theoretical total number of oil glands in the region, and the variance of the coordinate distribution.

[0087] (3)

[0088] In equation (3), The closer the value is to 1, the more intact the oil glands and the healthier the texture of the peel; the closer it is to 0, the more damaged and disordered the oil glands and the higher the degree of defects.

[0089] Step 2.5: Statistically analyze the grayscale values ​​of pixels in the local peel area and calculate the degree of grayscale dispersion.

[0090] (4)

[0091] In equation (4), This represents the grayscale value of a local pixel. The average gray level of the region. This is the texture grayscale fluctuation coefficient.

[0092] Step 2.6: Calculate the texture disorder coefficient TD based on the oil gland integrity index and texture grayscale fluctuation coefficient.

[0093] (5)

[0094] In formula (5), the larger TD is, the more severe the damage to the oil glands in the region and the more severe the gray disorder, and the higher the degree of the corresponding defect. It can accurately distinguish four states: healthy peel, slight epidermal wear, moderate defect, and severe damage.

[0095] Step 3: Calculate the defect irregularity based on the perimeter of the defect outline and the enclosed area. Perform normalization mapping on the defect irregularity to obtain the morphological regularity coefficient. Construct the mid-level morphological mechanism features using the defect irregularity and the morphological regularity coefficient. Among them, the defect irregularity is used to quantify the regularity of the defect, and the morphological regularity coefficient quantifies the morphological differences of the defect.

[0096] The specific steps are as follows:

[0097] Step 3.1: Perform edge detection and contour extraction on the marked defect area, remove fragmented noise contours by contour connectivity filtering, retain the complete outer contour of the defect target, and accurately solve the contour perimeter and enclosed area.

[0098] Step 3.2: Extract the perimeter P and the area A enclosed by the contour of a single defect target.

[0099] Step 3.3: Calculate the defect irregularity DI based on the perimeter of the contour and the area enclosed by the contour.

[0100] (6)

[0101] In equation (6), The closer the DI is to 1, the closer the shape is to a perfect circle (regular structure of ulcers, fruit stalks and fruit navels); the larger the DI is, the more disordered the outline (mechanical damage, irregular defects of mold).

[0102] Step 3.4: Perform normalization mapping on the defect irregularity to obtain the morphological regularity coefficient SI:

[0103] (7)

[0104] In equation (7), When SI approaches 1, the more regular the target morphology, the higher the probability of ulceration or intrinsic organ defects; when SI approaches 0, the more disordered the target morphology, the higher the probability of mechanical injury or mold defects. This enables refined quantitative differentiation of different defect morphologies.

[0105] This embodiment introduces defect irregularity to quantify the differentiated characteristics of defect shapes. For example, defects such as mechanical damage and mold have irregular shapes, while ulcers, fruit stalks, and fruit navels have relatively regular shapes. At the same time, a morphological regularity coefficient is introduced to further distinguish regular ulcer lesions, gradual mold changes, and irregular mechanical damage.

[0106] Step 4: Calculate the normalized distance of the region based on the geometric center coordinates of the fruit and the equivalent radius of the fruit, and solve the defect location factor based on the preset region weight factor, and solve the region fit factor according to the mechanism threshold discrimination rule; construct the high-level spatial distribution mechanism features with the defect location factor and the region fit factor; wherein, the defect location factor is used to quantify the probability of defects in different regions, and the region fit factor is used to determine whether the target region conforms to the inherent distribution law of defects.

[0107] The specific steps are as follows:

[0108] Step 4.1: Solve for the geometric center coordinates of the navel orange fruit by fitting the minimum circumcircle of the fruit. equivalent radius of fruit Determine the fruit region (equatorial region / stem region / navel region) where the defect is located based on the center coordinates of the defect target.

[0109] Step 4.2: Calculate the Euclidean distance from the defect target center (x, y) to the geometric center of the fruit:

[0110] (8)

[0111] In equation (8), d is the Euclidean distance.

[0112] Step 4.1: Calculate the normalized distance of the region based on the Euclidean distance and the equivalent radius of the fruit.

[0113] (9)

[0114] In equation (9), This represents the region-normalized distance.

[0115] Step 4.3: Assign regional weighting factors based on regional mechanisms: Equatorial region (high incidence of mechanical injuries) The fruit stem area (where ulcers are common) The fruit's blossom end area (where mold is most likely to occur) Low-incidence areas .

[0116] Step 4.4: Calculate the defect location factor PF based on the region weighting factor and the region normalized distance.

[0117] (10)

[0118] In equation (10), The regional weighting factor and the defect location factor are used to quantify the probability of defects occurring in different regions and suppress false detection results that do not conform to the distribution pattern.

[0119] Step 4.5: Based on the set mechanism threshold discrimination rules, solve for the region fitness factor MF:

[0120] (11)

[0121] In equation (11), The threshold value is 0.4 (determined by sample statistics and agricultural experience); MF=1 indicates that the current area conforms to the high incidence mechanism of defects and the enhancement feature is retained; MF=0.2 indicates that the area does not conform to the defect distribution pattern and is judged as a false defect area, the feature response is weakened, and false detection areas with contradictory mechanisms are filtered out.

[0122] Step 5: Calculate the fluorescence intensity index based on the average fluorescence grayscale values ​​of the background area, healthy fruit peel area, and suspected moldy area in the fluorescence image; calculate the area of ​​abnormal fluorescence response area and the total effective area of ​​the fruit to determine the proportion of infected area; calculate the mold infection degree coefficient based on the fluorescence intensity index and the proportion of infected area; construct the top-level latent mold mechanism characteristics using the fluorescence intensity index and the mold infection degree coefficient; whereby the fluorescence intensity index is used to quantify early mold defects, and the mold infection degree coefficient is used to quantify the differences between early latent infection, mid-term mycelial germination, and late-term large-area rot.

[0123] The specific steps are as follows:

[0124] Step 5.1: Standardize the ultraviolet fluorescence imaging light intensity and shooting distance, perform background grayscale correction on the fluorescence image, and eliminate environmental substrate fluorescence interference; extract the average fluorescence grayscale values ​​of the background area, healthy fruit peel area, and suspected moldy area respectively.

[0125] Step 5.2: Calculate the mean fluorescence gray value of the three types of regions: The mean value of background fluorescence in the image. The average fluorescence value of healthy fruit peel. The average fluorescence value is for areas suspected of mold growth.

[0126] Step 5.3: Calculate the normalized fluorescence intensity index FI:

[0127] (12)

[0128] In Equation (12), FI>1.5 indicates the presence of early moldy fluorescence abnormality in the region, and the larger the FI, the more severe the cell infection damage.

[0129] Step 5.4: Calculate the area of ​​the abnormal fluorescence response region. Total effective area of ​​fruit Solve for the percentage of infected area:

[0130] (13)

[0131] In equation (13), This represents the percentage of the area infested.

[0132] Step 5.5: Based on the fluorescence intensity index and the percentage of infected area, calculate the mold infection degree coefficient MI:

[0133] (14)

[0134] In Equation (14), MI simultaneously couples fluorescence response intensity and infection area to accurately quantify the differences in disease severity between early latent mold growth, mid-term mycelial spread, and late-term large-area rot.

[0135] In this embodiment, early mold defects lack obvious visible texture and exhibit extremely weak features in the RGB image, making them a major missed detection point for the YOLO11 model. Based on the ultraviolet fluorescence imaging mechanism, the damaged cell structure of the fruit peel in the mold-infected area will produce a specific fluorescence response, while healthy fruit peel will not show a fluorescence reaction. This embodiment introduces a fluorescence intensity index to quantify the characteristics of early mold defects; simultaneously, it introduces a mold infection degree coefficient to quantify the differences between early latent infection, mid-term mycelial germination, and late-term large-scale decay, adapting to the mold evolution pattern and achieving accurate identification and grading of early weak mold.

[0136] Step S300: The visual feature map is enhanced layer by layer using the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features to generate enhanced feature maps at each level. The enhanced feature maps at each level include: micro feature map, morphological optimization feature map, effective feature map, and cross-modal feature map.

[0137] In this embodiment, the YOLO11 model suffers from several industry-wide problems, including feature extraction lacking guidance from agricultural mechanisms, a disconnect between visual features and the pathological patterns of defects, poor adaptability of strong and weak defect features, severe confusion between early weak defects and peel texture features, and blind fusion of multi-scale defect features. This embodiment addresses these issues by performing quantified stratification on the multi-dimensional physiological parameters of navel oranges in step S200, obtaining multi-level mechanistic features composed of shallow texture mechanism features, mid-layer morphological mechanism features, high-layer spatial distribution mechanism features, and top-layer latent mold mechanism features.

[0138] Then, based on the hierarchical pathological differences in navel orange defects such as texture damage, morphological evolution, spatial distribution, and latent infection characterized by multi-level mechanism features, a four-level progressive mechanism fusion architecture is constructed, which includes "microscopic enhancement of shallow texture, precise depiction of morphology in the middle layer, global spatial constraint in the upper layer, and completion of latent mold features in the top layer". This achieves all-round targeted enhancement of defect features "from micro to macro, from explicit to implicit, and from local to global".

[0139] Specifically, the visual feature maps at multiple scales are enhanced layer by layer using shallow texture mechanism features, mid-layer morphological mechanism features, high-layer spatial distribution mechanism features, and top-layer latent mold mechanism features, including:

[0140] Step S301: Based on the shallow texture mechanism features, perform texture micro-enhancement on the visual feature map to generate a micro-feature map.

[0141] In this embodiment, the shallow layer is a micro-feature activation layer. This layer addresses the problem of missed detection caused by early defects in navel oranges, such as the lack of clear outlines, weak textures, and easy blending with normal peel textures. This layer focuses on low-discrimination targets such as minor epidermal abrasion, early latent mold, and shallow micro-ulcers. Starting from the microscopic oil gland texture mechanism of the peel, it achieves pixel-level enhancement of weak defect features, providing an effective micro-feature base for subsequent morphological and spatial layer fusion. This is a fundamental prerequisite for global fine feature enhancement.

[0142] In this embodiment, the oil glands of healthy navel orange peel are evenly and densely distributed in a dotted pattern, with a regular and orderly texture; defective areas show characteristics such as oil gland rupture, necrosis, disordered distribution, and uneven grayness. Moreover, the degree of defect incidence is positively correlated with the degree of texture disorder and the degree of oil gland damage, which is the core microscopic mechanism for distinguishing healthy peel from early weak defects.

[0143] Therefore, the specific method for performing texture micro-enhancement on visual feature maps is as follows:

[0144] First, texture enhancement weights are constructed by fusing the oil gland integrity index (OGI) and the texture disorder coefficient (TD). :

[0145] (15)

[0146] In equation (15), For the sigmoid function, The higher the texture disorder and the lower the integrity of the oil glands in the defect area, the greater the texture enhancement weight, thus achieving adaptive lifting of the micro-region of the defect.

[0147] Then, texture micro-enhancement is performed on the visual feature map using texture enhancement weights:

[0148] (16)

[0149] In the formula, It is a multi-scale visual feature map, which includes the basic texture of the fruit peel and edge pixel information, without any mechanistic constraints; This is pixel-level element-wise multiplication. This is a microscopic feature map after texture enhancement.

[0150] The microscopic feature map in this embodiment can effectively amplify the texture differences between early defects and normal peel, and suppress interference from ineffective textures in healthy peel.

[0151] Step S302: Based on the mid-level morphological mechanism features, perform differentiated morphological enhancement on the micro-feature map to generate a morphologically optimized feature map.

[0152] In this embodiment, the middle layer is a visible feature optimization layer. It inherits the micro-texture features enhanced by the shallow layer, completes the morphological differentiation of molding defects, solves the problem of homogenization of different types of defects and low feature differentiation, accurately shapes the exclusive morphological features of ulcers, mechanical damage, mildew, and sunburn, and connects micro-texture with macro-spatial features.

[0153] In this embodiment, the various defect morphologies of navel oranges have strong specificity: ulcer lesions are regular in shape, nearly round, and have clear boundaries; mechanical damage and mold are disordered in shape, diffuse at the edges, and have no fixed geometric features; the morphology of the fruit stalk and navel is stable and regular, and there is an essential difference from the defect morphology. Differentiated features can be enhanced through morphological parameters.

[0154] Therefore, the specific method for performing differential morphological enhancement on the micro-feature map is as follows:

[0155] Based on Defect Irregularity Derivative (DI), a Morphological Adaptive Enhancement Weight is Constructed To achieve differentiated enhancement for different defect morphologies:

[0156] (17)

[0157] A two-way enhancement logic is applied to both regular morphologies (ulcers, intrinsic organs) and disordered morphologies (mechanical injuries, fungal infections) to output morphology-optimized feature maps. :

[0158] (18)

[0159] In equation (18), This is the morphological balance coefficient; When the value approaches 1, the ulcerative lesions exhibit enhanced characteristics of raised edges and central depression. As the value approaches zero, the diffuse and blurred morphological features of mechanical damage and mold are enhanced, achieving differentiated characterization of "the boundary of the enhanced morphological features of regular defects and the region of the enhanced morphological features of disordered defects".

[0160] Step S303: Based on the high-level spatial distribution mechanism characteristics, perform pseudo-defect filtering on the morphological optimization feature map to generate an effective feature map.

[0161] In this embodiment, the high layer is the global feature constraint layer. Based on the agricultural mechanism of fixed spatial distribution of defects in navel oranges, the texture and morphological features optimized in the previous step are globally screened and weighted to filter out pseudo-defect features that do not conform to the distribution pattern, strengthen the effective features in high-defect areas, and achieve the mechanism matching between local features and global fruit structure.

[0162] In this embodiment, the spatial distribution of defects in navel oranges exhibits strong regularity: mechanical injuries are common on the equatorial side, ulcers are common around the fruit stem, mold is common in the sunken area of ​​the fruit navel, and the incidence of defects on the shaded side and central area of ​​the fruit is extremely low. Suspected features that do not conform to this distribution pattern are mostly fruit peel texture noise and light and shadow interference pseudo-features.

[0163] Therefore, the specific method for performing pseudo-defect filtering on the morphological optimization feature map is as follows:

[0164] First, spatial adaptive weights are constructed by fusing the defect location factor PF and the region fit factor MF. :

[0165] (19)

[0166] In equation (19), The defect location factor is the spatial defect location factor of the target area, representing the probability of the corresponding defect occurring in the current area; This is a regional fit factor used to determine whether a region matches attributes with high incidence of defects.

[0167] Then, the features are optimized in the spatial dimension to output an effective feature map. :

[0168] (20)

[0169] In equation (20), when the target is located in a high-defect area, Approaching 1, retaining and enhancing effective defect characteristics; when the target is in a low-defect area, Approaching 0.2 weakens pseudo-defect features, completing feature purification and correction from a global spatial dimension.

[0170] Step S304: Based on the top-level latent mold growth mechanism features, perform latent defect enhancement on the effective feature map to generate a cross-modal feature map.

[0171] The top layer of this embodiment is a dedicated latent defect completion layer, which supplements the dimensions of the first three levels of visible light feature system. It specifically solves the problem of missing latent defects in early mold growth in RGB images due to the lack of visual features and the inability to identify them in the first three levels, thus achieving full-dimensional feature coverage of both visible and latent defects.

[0172] In this embodiment, early mold growth only occurs within the pericarp cells, with no visible texture or morphological changes on the epidermis, and no characteristic response in visible light images; however, damaged mold cells will produce a specific fluorescence enhancement response under ultraviolet light, and the fluorescence intensity is positively correlated with the degree of infection, which is the only effective mechanistic feature for identifying early latent mold growth.

[0173] Therefore, the specific method for performing latent defect enhancement on the effective feature map is as follows:

[0174] First, a latent feature activation weight is constructed based on the fluorescence intensity index FI and the mold infection degree coefficient MI. : (twenty one);

[0175] Then, cross-modal feature fusion and completion are performed to output cross-modal feature maps. :

[0176] (twenty two)

[0177] In equation (22), Feature-based channel-by-channel overlay and fusion; The activation weights for latent mold features are determined by the degree of infection and the significance of fluorescence response. The higher the degree of infection and the more significant the fluorescence response, the greater the weight. Feature completion and activation are performed on latent mold regions that do not respond to visible light, thereby achieving effective extraction of early mold features.

[0178] Step S400: Perform adaptive mechanism global fusion on the enhanced feature maps of each level and the visual feature map to generate a mechanism-enhanced hybrid feature map, and obtain the mechanism-enhanced hybrid feature map.

[0179] Mechanism-enhanced hybrid feature map of this embodiment The calculation formula is:

[0180] (twenty three)

[0181] In the formula, As a learnable and adaptive weight coefficient, the contribution ratio of each mechanism dimension is dynamically adjusted according to different defect types and different severity of disease, so as to achieve on-demand enhancement and precise adaptation, and solve the problem of blindness in pure data-driven feature extraction.

[0182] Therefore, this embodiment uses the multi-scale visual feature map output by the neck network and a complete set of quantized mechanism parameters such as OGI, TD, SI, MF, MI, and FI as prior constraints to perform differentiated weighted enhancement on defect features of different developmental stages, types, and scales. This specifically compensates for the shortcomings of the YOLO11 lightweight network in terms of weak feature representation ability and insufficient extraction of weak defects. Finally, it outputs a hybrid enhanced feature map with high discriminativeness and strong mechanism correlation, providing a high-quality and highly specific pre-feature base for the subsequent interference suppression module to distinguish between true and false defects and to purify features.

[0183] Step S500: Construct an adaptive suppression constraint term for the navel orange interference target, and use the adaptive suppression constraint term to perform mechanistic constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map.

[0184] In this embodiment, after the processing in step S300, suspected defect features are enhanced across the entire domain. However, while amplifying the actual defect features, interfering features that are visually similar to the defect, such as the fruit stalk and navel, are also amplified simultaneously, resulting in feature confusion in the feature map. To address this, this embodiment utilizes the fundamentally different mechanisms between defects and the inherent organs of the navel orange to precisely screen and flexibly suppress the enhanced mixed features.

[0185] Specifically, the adaptive inhibition constraint term includes a morphological mechanism constraint term created by the morphological regularity coefficient and a spatial distribution mechanism constraint term created by the region fit factor; the morphological mechanism constraint term is used to perform inhibition on the region where the morphologically regular interference target is located, and the spatial distribution mechanism constraint term is used to perform inhibition on the region that conforms to the inherent organ distribution law.

[0186] Therefore, the specific implementation of morphological and spatial mechanistic constraints on the hybrid feature map enhanced by the adaptive suppression constraint term is as follows: extract the pixel-level binary mask of the interfering target; calculate the adaptive suppression weight coefficient based on the pixel-level binary mask, spatial distribution mechanistic constraint term, and morphological mechanistic constraint term; perform mechanistic flexible feature purification on the hybrid feature map enhanced by the mechanistic constraint term according to the adaptive suppression weight coefficient, generate the purified defect feature map, and use the purified defect feature map as the final feature map.

[0187] The main interference targets include the fruit stalk and the fruit navel. Therefore, the mechanism discrimination mask for the fruit stalk / fruit navel interference region is:

[0188] (twenty four)

[0189] In equation (24), A pixel-level binary mask is used to accurately mark the regions in the image that belong to the inherent interfering organs of navel oranges, providing a basis for regional localization for subsequent flexible suppression.

[0190] The adaptive suppression weight coefficient is as follows:

[0191] (25)

[0192] In equation (25), The dynamic suppression weight takes values ​​in the range [0,1]. To achieve morphological mechanism constraints, the inhibition weight is increased for well-formed, inherent organ regions, and the inhibition weight is decreased for morphologically disordered, defective regions. To achieve spatial distribution mechanism constraints, regions that conform to the inherent organ distribution pattern are strengthened and suppressed, while regions with high incidence of matching defects are weakened and suppressed. This achieves an adaptive adjustment logic where the closer the mechanism is to the interference, the stronger the suppression.

[0193] Among them, the defect feature map is purified and output from the mechanism flexible feature. for:

[0194] (26)

[0195] In the formula, The mechanism-enhanced hybrid feature map is output from the hierarchical fusion of multi-level mechanism features in step S300. This is a pixel-by-pixel multiplication operation.

[0196] This embodiment dynamically adjusts the inhibition intensity by using mechanistic weights. While completely filtering out the interference features of fruit stalks and fruit navels that are mistakenly enhanced, it fully preserves the effective features of early weak defects and irregular defects, thus solving the dual problems of feature confusion and over-suppression.

[0197] Step S600: Input the final feature map into the head network of the YOLO11 model, and perform defect detection using the shared convolution strategy to obtain the defect detection result.

[0198] In this embodiment, as Figure 2 As shown, the head network sequentially includes 3 parallel Conv-GN modules, 2 cascaded Conv-GN modules, 3 Conv-reg modules, 3 Conv-cls modules, and 3 Scale modules. Each Conv-reg module is cascaded with one Scale module. The cascaded structure of the 3 parallel Conv-reg modules and the Scale modules constitutes the regression branch. The 3 Conv-cls modules are in parallel to form the classification branch. Both the classification branch and the regression branch are connected to the 2 cascaded Conv-GN modules.

[0199] The shared convolution strategy of the head network is as follows: three parallel Conv-GN modules are used to perform independent 3×3 convolution group normalization preprocessing on the three feature maps of different scales output by the neck network to obtain three sets of preprocessed features. The three sets of preprocessed features are then input into two cascaded Conv-GN modules with completely shared weights to perform cross-scale feature sharing fusion and obtain shared fused features. The first Conv-GN module (which is connected to the three parallel Conv-GN modules) is a 3×3 shared convolution group normalization, and the second Conv-GN module is a 1×1 shared convolution group normalization.

[0200] The shared fusion features are input to the classification branch and the regression branch respectively. The classification branch outputs the category prediction result through 1×1 convolution with shared weights across all branches. The regression branch outputs the initial coordinates and size prediction through 1×1 convolution with shared weights across all branches. The scale of the size prediction is adaptively corrected based on the scaling module to obtain the corrected size prediction. Non-maximum suppression processing is performed on the category prediction result, the initial coordinates, and the corrected size prediction to generate the defect detection result.

[0201] Therefore, the head network utilizes a shared convolution strategy in deep learning, applying the same weight parameters at different locations within the network to process the final feature maps at three different scales output by the neck network. The outputs are then connected to a scaling module to scale the images proportionally to the size of the targets detected by each head. This improves the utilization of semantic information in the network and effectively balances the model's detection accuracy with the requirements of lightweight deployment.

[0202] Therefore, this embodiment quantifies the multi-dimensional physiological parameters of the navel orange's physiological structure to distinguish the differentiated characteristics of different defects and different degrees of disease, providing prior support for the accurate identification of navel orange defects and improving the accuracy of defect identification. It adopts a hierarchical enhancement strategy to adaptively match the feature requirements of defects of different scales and developmental stages, solving the problem of insufficient feature extraction and serious feature confusion of the YOLO11 model for weak defects, small-sized defects, and irregular defects. By constructing adaptive suppression constraints, it performs morphological and spatial mechanistic constraints on the hybrid feature map of mechanism enhancement, which can accurately eliminate the false detection interference of interfering targets such as fruit stalks, navels, and natural textures.

[0203] Example 2

[0204] Figure 3 This is a block diagram of a navel orange defect detection system provided in one embodiment of the present invention. Figure 3 As shown, this embodiment provides a navel orange defect detection system for implementing the navel orange defect detection method in Embodiment 1. The system includes:

[0205] The feature acquisition module is used to obtain the multi-scale visual feature map output by the neck network of the YOLO11 model after the YOLO11 model performs multi-scale feature extraction on the navel orange image.

[0206] The mechanism quantification module is used to extract multi-dimensional physiological parameters of navel oranges and perform quantification stratification to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features.

[0207] The feature enhancement module is used to sequentially enhance the visual feature map by utilizing the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features, generating enhanced feature maps at each level.

[0208] The feature fusion module is used to perform adaptive global fusion of enhanced feature maps at each level with visual feature maps to generate a hybrid feature map with enhanced mechanism.

[0209] The mechanism constraint module is used to construct an adaptive suppression constraint term for the navel orange interference target. The adaptive suppression constraint term is used to perform mechanism constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map.

[0210] The defect detection module is used to input the final feature map into the head network of the YOLO11 model, perform defect detection using a shared convolution strategy, and obtain the defect detection results.

[0211] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the navel orange defect detection method in Embodiment 1.

[0212] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the navel orange defect detection method in Embodiment 1.

[0213] This embodiment quantifies the multi-dimensional physiological parameters of the navel orange's physiological structure to obtain mechanistic quantitative features, thereby distinguishing the differentiated characteristic patterns of different defects and different degrees of disease. This provides prior support for the accurate identification of navel orange defects and improves the accuracy of defect recognition. A hierarchical enhancement strategy is adopted to adaptively match the feature requirements of defects at different scales and developmental stages, solving the problem of insufficient feature extraction and severe feature confusion of the YOLO11 model for weak defects, small-sized defects, and irregular defects. By constructing adaptive suppression constraints, the morphological and spatial mechanistic constraints on the mechanistic enhancement hybrid feature map can be applied to accurately eliminate false detection interference from interfering targets such as fruit stalks, navels, and natural textures.

[0214] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0215] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0216] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for detecting defects in navel oranges, characterized in that, The method includes: After performing multi-scale feature extraction on the navel orange image using the YOLO11 model, the multi-scale visual feature map output by the neck network of the YOLO11 model is obtained. Multidimensional physiological parameters of navel oranges were extracted and quantitative stratification was performed to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features. By utilizing the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features, the visual feature map is enhanced layer by layer to generate enhanced feature maps at each level. Adaptive mechanism-based global fusion is performed between the enhanced feature maps at each level and the visual feature maps to generate a hybrid feature map with enhanced mechanism. An adaptive suppression constraint term for the navel orange interference target is constructed. The adaptive suppression constraint term is used to perform mechanistic constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map. The final feature map is input into the head network of the YOLO11 model, and defect detection is performed using a shared convolution strategy to obtain the defect detection results.

2. The method for detecting defects in navel oranges according to claim 1, characterized in that, The YOLO11 model includes a backbone network, a neck network, and a head network; The backbone network adopts the StarNet structure to perform multi-scale feature extraction on the navel orange image, resulting in four initial feature maps at different scales. The neck network adopts a BiFPN structure to perform different scale fusion and nonlinear enhancement on the initial feature maps of four different scales to obtain fused features of three scales, and uses the fused features of three scales as the visual feature map. The head network is used to perform defect detection on the final feature map using a shared convolution strategy to obtain defect detection results.

3. The method for detecting defects in navel oranges according to claim 2, characterized in that, The backbone network has parameterless attention units embedded at its ends, which are used to perform attention enhancement processing on the initial feature map at the smallest scale.

4. The method for detecting defects in navel oranges according to claim 1, characterized in that, Multidimensional physiological parameters of navel oranges were extracted and quantitative stratification was performed, including: The effective number of oil glands, total number of oil glands, perimeter and enclosed area of ​​defects, geometric center coordinates of the fruit and equivalent radius of the fruit, and the mean fluorescence gray value of the background area, healthy peel area and suspected moldy area in the fluorescence image were collected. Based on the number of effective oil glands and the total number of oil glands, the oil gland integrity index is calculated, the gray-scale dispersion of local peel areas is statistically analyzed, and the texture disorder coefficient is calculated based on the oil gland integrity index and the gray-scale dispersion. The shallow texture mechanism characteristics are constructed using the oil gland integrity index and the texture disorder coefficient. Based on the perimeter of the defect's outline and the area it encloses, the defect irregularity is calculated. A normalization mapping is then performed on the defect irregularity to obtain the morphological regularity coefficient. The mid-layer morphological mechanism features are then constructed using the defect irregularity and the morphological regularity coefficient. Based on the geometric center coordinates of the fruit and the equivalent radius of the fruit, the normalized distance of the region is calculated, and the defect location factor is solved based on the preset region weight factor. The region fitness factor is solved according to the mechanism threshold discrimination rule. The high-level spatial distribution mechanism features are constructed using the defect location factor and the region fitness factor. The fluorescence intensity index is calculated based on the average fluorescence grayscale values ​​of the background area, healthy peel area, and suspected moldy area in the fluorescence image; the area of ​​abnormal fluorescence response area and the total effective area of ​​the fruit are statistically analyzed to determine the proportion of infected area; the mold infection degree coefficient is determined based on the fluorescence intensity index and the proportion of infected area; and the top-level latent mold mechanism characteristics are constructed using the fluorescence intensity index and the mold infection degree coefficient.

5. The method for detecting defects in navel oranges according to claim 4, characterized in that, The enhanced feature maps at each level include: microscopic feature maps, morphological optimization feature maps, effective feature maps, and cross-modal feature maps; multi-scale visual feature maps are sequentially enhanced using shallow texture mechanism features, mid-level morphological mechanism features, high-level spatial distribution mechanism features, and top-level latent mold mechanism features, including: Based on the shallow texture mechanism features, micro-texture enhancement is performed on the visual feature map to generate a micro-feature map; Based on the mid-level morphological mechanism characteristics, differential morphological enhancement is performed on the micro-feature map to generate a morphologically optimized feature map; Based on the spatial distribution mechanism characteristics of high-level structures, pseudo-defect filtering is performed on the morphological optimization feature map to generate an effective feature map. Based on the top-level latent mold growth mechanism characteristics, latent defect enhancement is performed on the effective feature map to generate cross-modal feature maps.

6. The method for detecting defects in navel oranges according to claim 4, characterized in that, The adaptive inhibition constraint term includes a morphological mechanism constraint term created by the morphological regularity coefficient and a spatial distribution mechanism constraint term created by the region fit factor; the morphological mechanism constraint term is used to perform inhibition on the region where the morphologically regular interference target is located, and the spatial distribution mechanism constraint term is used to perform inhibition on the region that conforms to the inherent organ distribution law.

7. The method for detecting defects in navel oranges according to claim 6, characterized in that, Mechanistic constraints on the execution morphology and space of hybrid feature maps based on constraint terms for mechanism enhancement include: Extract the pixel-level binary mask of the interfering target; Based on pixel-level binary masks, spatial distribution mechanism constraints, and morphological mechanism constraints, the computer adaptive suppression weight coefficients are calculated. Based on the mechanism-adaptive suppression weight coefficient, the mechanism-enhanced hybrid feature map is subjected to mechanism-flexible feature purification to generate a purified defect feature map, which is then used as the final feature map.

8. A navel orange defect detection system, used to implement the navel orange defect detection method according to any one of claims 1-7, characterized in that, The system includes: The feature acquisition module is used to obtain the multi-scale visual feature map output by the neck network of the YOLO11 model after the YOLO11 model performs multi-scale feature extraction on the navel orange image. The mechanism quantification module is used to extract multi-dimensional physiological parameters of navel oranges and perform quantification stratification to obtain shallow texture mechanism features, middle morphological mechanism features, high spatial distribution mechanism features, and top latent mold mechanism features. The feature enhancement module is used to sequentially enhance the visual feature map by utilizing the shallow texture mechanism features, the middle morphological mechanism features, the high spatial distribution mechanism features, and the top latent mold mechanism features, generating enhanced feature maps at each level. The feature fusion module is used to perform adaptive global fusion of enhanced feature maps at each level with visual feature maps to generate a hybrid feature map with enhanced mechanism. The mechanism constraint module is used to construct an adaptive suppression constraint term for the navel orange interference target. The adaptive suppression constraint term is used to perform mechanism constraints on the hybrid feature map of mechanism enhancement in the space and morphology of the navel orange to obtain the final feature map. The defect detection module is used to input the final feature map into the head network of the YOLO11 model, perform defect detection using a shared convolution strategy, and obtain the defect detection results.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method for detecting defects in navel oranges according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for detecting defects in navel oranges as described in any one of claims 1-7.

Images