Edge computing system for hyperspectral maturity detection of apples

Abstract

The present application relates to the technical field of edge computing, in particular to an edge computing system for apple hyperspectral maturity detection. The system obtains apple hyperspectral spectral cubes through a hyperspectral data receiving interface, extracts spectral segment reflection, spectral shape change, local absorption and continuous spectral envelope shift features by an edge computing processor, and constructs a maturity sensitive spectral segment atlas with maturity label contribution, spectral segment redundancy and edge computing cost to determine a basic spectral segment subset; meanwhile, the system trains an edge student model with knowledge constraint by a full spectrum teacher model, and triggers a supplementary spectral segment reasoning branch according to grade confidence interval and spectral shape residual during edge reasoning, and outputs maturity grade after fusing the basic reasoning result and the supplementary reasoning result. The system can reduce the load of full spectrum data participating in edge computing, and retain spectral discrimination information of adjacent maturity grades.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, specifically to an edge computing system for detecting the hyperspectral maturity of apples. Background Technology

[0002] Hyperspectral analysis of apple maturity typically utilizes the reflectance response of apples across multiple consecutive wavelengths to form spectral cube data. A discriminant model is then established based on spectral differences caused by variations in peel tissue, moisture, sugar-acid conversion, and internal substances at different maturity stages. Conventional methods generally involve first performing black-and-white correction, noise smoothing, normalization, and background removal on the acquired hyperspectral data. Then, reflectance intensity, absorption peak positions, spectral slope, or several statistical features are extracted from the complete spectral curve. Finally, a classification or regression model is used to output the maturity level. Because the complete spectral cube contains both spatial and spectral dimensions, the data volume is large. Conventional detection processes are mostly completed on high-performance upper-level computing devices, with edge computing only handling data reception, preprocessing, or result forwarding. While this approach can establish maturity discrimination relationships using full-spectrum information, continuous on-site detection requires caching the complete spectral cube and performing numerous spectral calculations. Data reading, feature generation, and model inference all consume significant edge computing resources.

[0003] To adapt to edge computing conditions, existing technologies typically employ fixed spectral band selection, principal component analysis (PCA) dimensionality reduction, lightweight classification models, or model compression to reduce computational load. Fixed spectral band selection generally selects several feature bands based on the correlation between a single spectral band and maturity labels in historical samples, and then inputs the selected bands into the edge model for inference. PCA dimensionality reduction typically projects the full-spectrum data into a small number of integrated features to reduce input dimensionality. Lightweight models reduce the inference load by decreasing the number of network layers, nodes, or feature channels. A common feature of these approaches is that they treat spectral band compression and model compression as relatively independent processing steps. Spectral band selection focuses primarily on historical statistical contributions, while model compression focuses primarily on computational scale. The edge inference process typically uses fixed inputs and fixed paths. For samples with significant maturity differences, these approaches can yield usable results. However, for samples approaching maturity, moderate maturity, and overmaturity, the complementary relationships, weak absorption variations, and local spectral residuals between a few sensitive spectral bands are easily weakened during dimensionality reduction or compression.

[0004] The main technical problem with existing edge computing solutions for apple hyperspectral maturity detection lies in the direct conflict between full-spectrum detection and limited computing resources at the edge. Conventional fixed-segment selection or simple model compression struggles to retain the subtle spectral discrimination information required for adjacent maturity levels. This is because apple maturity differences are not uniformly distributed across the entire spectrum, but rather manifest in certain sensitive segments and their combinations. If selection is based solely on the contribution of a single spectral segment, segments that supplement maturity boundaries are easily lost. Simply reducing the model size is insufficient to preserve the level boundaries, sample similarity relationships, and spectral details formed under full-spectrum conditions. Furthermore, if the edge uses the same inference path for all samples, it cannot distinguish between easily discriminative samples and boundary samples, leading to both ineffective computation and boundary misjudgments. Therefore, a closed-loop synergistic technical solution is needed that can achieve this between spectral segment selection, model compression, and edge inference triggering. Summary of the Invention

[0005] The purpose of this invention is to provide an edge computing system for detecting the hyperspectral maturity of apples, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An edge computing system for detecting the hyperspectral maturity of apples includes a hyperspectral data receiving interface, an edge computing processor, and a memory, wherein the memory stores instructions that can be executed by the edge computing processor. The instruction causes the edge computing processor to extract spectral reflectance features, spectral shape change features, local absorption features, and continuous spectrum envelope shift features from the apple hyperspectral cubic array, and to form a maturity-sensitive spectral map with spectral segments as nodes and using maturity label contribution, spectral segment redundancy, and edge computing cost. Based on the maturity-sensitive spectral segment map, a basic spectral segment subset is determined. Based on the full-spectrum teacher model, a marginal student model that only inputs the basic spectral segment subset is trained with knowledge constraints. During marginal inference, a supplementary spectral segment inference branch is triggered based on the level confidence interval and spectral residual. The maturity level is output after the basic inference result and the supplementary spectral segment inference result are consistently fused.

[0007] Preferably, the construction of the maturity-sensitive spectral segment map includes: calculating the discriminant contribution value between each spectral segment in the apple hyperspectral cube and the maturity level label; For any two spectral bands, calculate the redundant correlation value based on spectral shape correlation and local absorption position similarity, and generate the computational cost based on the amount of data read, cache usage, and multiply-accumulate operations when the spectral bands participate in edge calculation; The discrimination contribution value, the redundancy correlation value, and the computational cost value are written into the same spectral weight expression, so that the spectral segment selection is simultaneously constrained by maturity discrimination information, spectral segment complementarity relationship, and edge computing load.

[0008] Preferably, the knowledge constraint training of the full-spectrum teacher model on the marginal student model includes: using the maturity level probability distribution obtained under full-spectrum input, the embedding distance of adjacent maturity level boundary samples, and the maturity similarity matrix of samples in the same batch as constraint objects; The rank probability, feature embedding, and sample similarity matrix obtained by the marginal student model under the input of the basic spectral segment subset are respectively constrained with the constraint object, and an adjacent maturity rank boundary preservation term is added to the training loss; The marginal student model retains the distinguishing boundaries between near-mature, moderately mature, and over-mature samples under full-spectrum conditions.

[0009] Preferably, the triggering of the supplementary spectral segment inference branch includes: after the marginal student model outputs the basic inference result, calculating the level confidence interval between the highest candidate level of maturity and the second highest candidate level; Calculate the spectral shape residual after the basic spectral segment subset is reconstructed to the full spectral reference space, and determine the sensitive spectral segment missing risk value based on the adjacency contribution of the spectral segments that did not participate in the inference in the maturity sensitive spectral segment map; When the confidence interval, the spectral residual, and the risk value of missing sensitive spectral segments meet the joint triggering condition, a supplementary spectral segment associated with the spectral residual is selected for secondary inference.

[0010] Preferably, the determination of the basic spectral segment subset includes: first selecting candidate spectral segments from the maturity-sensitive spectral segment map whose discrimination contribution values ​​meet the maturity differentiation requirements; Then, mutual exclusion screening is performed on the candidate spectral segments based on the redundant correlation values, and complementary compensation scores are calculated for the adjacent nodes of the excluded spectral segments. If there is a maturity boundary complementary relationship between the excluded spectral segments and the retained spectral segments, the adjacent spectral segments with low computational cost values ​​are added to the candidate spectral segments. Finally, the candidate spectral segments are combined and verified according to the maturity level coverage integrity, so that the basic spectral segment subset simultaneously covers the spectral discrimination intervals of immature, near-mature, moderately mature, and overmature.

[0011] Preferably, the marginal student model includes a spectral band combination coding layer, a maturity boundary embedding layer, and a lightweight classification layer; The spectral band combination coding layer receives the basic spectral band subset and the spectral band differential combination features constructed from the basic spectral band subset; The maturity boundary embedding layer constrains the interval between adjacent maturity level samples based on the boundary sample embedding distance output by the full-spectrum teacher model. The lightweight classification layer forms maturity level probabilities based on the low-dimensional features output by the maturity boundary embedding layer, so that the trained marginal student model maintains a fixed input mapping relationship with the basic spectral segment subset.

[0012] Preferably, the selection of the supplementary spectral band includes: based on the position of the residual peak in the full-spectrum reference space of the spectral residual; From the maturity-sensitive spectral segment map, retrieve spectral segment nodes that are adjacent to the position of the residual peak and have a maturity boundary contribution, sort the retrieved spectral segment nodes according to the redundancy correlation value, and then supplement the number of spectral segments according to the calculation cost limit. The sorted supplementary spectral segments are input into the supplementary spectral segment inference branch, which is independent of the basic inference path, so that the secondary inference only uses the local spectral segment information corresponding to the spectral anomaly of the current apple sample.

[0013] Preferably, the consistency fusion includes: jointly encoding the maturity level probability in the basic inference result, the maturity level probability output by the supplementary spectral segment inference branch, and the boundary contribution weight of the supplementary spectral segment in the maturity sensitive spectral segment map to form a basic evidence vector and a supplementary evidence vector. When the basic evidence vector and the supplementary evidence vector point to the same maturity level, the level probability is updated according to the boundary contribution weight. When both point to adjacent maturity levels, the embedding distance corresponding to the adjacent level boundary preservation item is called for re-discrimination and generation of fusion maturity level.

[0014] Preferably, the combined verification of the maturity level coverage integrity includes: mapping the basic spectrum subset to a continuous maturity axis from immature to overmature, calculating the intraclass dispersion and adjacent level interval of the samples corresponding to each maturity level under the basic spectrum subset, and matching the calculation results with the adjacent level boundary embedding distance formed by the full spectrum teacher model under the full spectrum input; When the intraclass dispersion of any maturity level exceeds the interval between its adjacent levels, a spectral node with a boundary complementary relationship and low computational cost is backtracked from the maturity-sensitive spectral segment map and added to the basic spectral segment subset.

[0015] Preferably, the generation of the fusion maturity level includes: when the basic evidence vector and the supplementary evidence vector point to adjacent maturity levels, extracting the maturity boundary embedding layer output features of the marginal student model, projecting the intermediate features of the supplementary spectral inference branch to the same embedding space, and calculating the distances of the two types of features to the adjacent maturity level boundary prototypes respectively. A boundary correction coefficient is generated based on the distance difference and the boundary contribution weight, and the fusion maturity level is determined by redistributing the probabilities of adjacent maturity levels using the boundary correction coefficient.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using a maturity-sensitive spectral segment map, the contribution of maturity labels, spectral segment redundancy, and edge computing costs are uniformly expressed. This ensures that the determination of the basic spectral segment subset no longer relies solely on the correlation between a single spectral segment and a maturity label, but also considers the complementary relationships between spectral segments and the computational load at the edge. Furthermore, a full-spectrum teacher model is used to constrain the edge student model with probability distribution, sample similarity relationships, and adjacent maturity level boundaries. This allows the edge student model, which only receives a basic spectral segment subset as input, to still retain the maturity discrimination boundary formed under full-spectrum conditions. Thus, the system reduces the involvement of full-spectrum data in computation while retaining key distinguishing criteria between near-mature, moderately mature, and overmature samples, alleviating the contradiction between excessive resource consumption by full-spectrum inference and the loss of boundary information caused by fixed dimensionality reduction.

[0017] 2. A joint triggering mechanism based on the confidence interval, spectral residual, and risk value of missing sensitive spectral segments is introduced at the edge inference stage. This mechanism initiates a supplementary spectral segment inference branch for maturity boundary samples or samples with anomalous spectral shapes, and then fuses the basic inference results with the supplementary spectral segment inference results for consistency. This approach allows samples with clear maturity characteristics to complete inference using a subset of basic spectral segments, reducing unnecessary supplementary calculations. For samples that are difficult to fully explain using basic spectral segments, relevant supplementary spectral segments are selected from the maturity sensitive spectral segment map based on the spectral residual for secondary discrimination. Basic inference, supplementary inference, and consistency fusion together form a processing path that allocates computational resources according to the sample discrimination difficulty, controlling edge-end cache usage, spectral segment reading volume, and inference load, while enhancing the consistency of maturity output for boundary samples. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall processing flow of an edge computing system for detecting the hyperspectral maturity of apples. Figure 2 Flowchart for constructing maturity-sensitive spectral segments and determining the basic spectral segment subset; Figure 3 Flowchart for training the knowledge constraint model of the full-spectrum teacher model on the marginal student model; Figure 4 To supplement the flowchart of spectral segment inference triggering and consistency fusion. Detailed Implementation

[0019] In one embodiment, reference Figure 1An edge computing system for detecting the hyperspectral maturity of apples includes a hyperspectral data receiving interface, an edge computing processor, and a memory. The memory stores instructions executable by the edge computing processor. These instructions enable the edge computing processor to extract spectral reflectance features, spectral shape variation features, local absorption features, and continuous spectrum envelope shift features from the apple hyperspectral spectral cube. Using spectral segments as nodes and incorporating maturity label contribution, spectral redundancy, and edge computing costs, a maturity-sensitive spectral segment map is formed. A basic spectral segment subset is determined based on this map. An edge student model, receiving only the basic spectral segment subset as input, is trained with knowledge constraints based on a full-spectrum teacher model. During edge inference, supplementary spectral segment inference branches are triggered based on the level confidence interval and spectral shape residuals. The maturity level is output after consistent fusion of the basic and supplementary spectral segment inference results. Specifically,... During implementation, the hyperspectral data receiving interface is used to receive the apple hyperspectral spectral cube that has completed basic correction. The edge computing processor reads the reflection sequence according to the spectral segment dimension and generates data representations related to maturity discrimination for each spectral segment. The memory stores the maturity-sensitive spectral segment map, edge student model parameters, constraint information generated by the full-spectrum teacher model, and running instructions for the supplementary spectral segment inference branch. During detection, the edge computing processor does not directly perform unified inference on all spectral segments. Instead, it determines the basic spectral segment subset participating in basic inference based on the maturity-sensitive spectral segment map. When the basic inference result is insufficient to stably distinguish adjacent maturity levels, it calls the supplementary spectral segment inference branch. In this embodiment, spectral segment selection, model constraints, and edge inference triggering form a closed processing chain, so that apple maturity detection avoids the direct conflict between full-spectrum unified calculation and fixed dimensionality reduction losing boundary information in the edge computing environment.

[0020] In this embodiment, the spectral reflectance characteristics are the mean reflectance, reflectance distribution dispersion, and reflectance variation range of a single spectral segment or a group of continuous spectral segments within the effective area of ​​the fruit surface. The spectral shape variation characteristics are the reflectance difference between adjacent spectral segments, local slope changes, and the degree of spectral shape curvature. The local absorption characteristics are the depth, width, and absorption position shift of the absorption valley formed under the preset continuous spectral envelope. The continuous spectral envelope shift characteristics are the upward, downward, or local deviation of the sample spectral curve relative to the maturity reference envelope. The edge computing processor establishes a unified index for the above characteristics according to the spectral segment number and links the maturity label contribution of each spectral segment with its corresponding index. The redundancy between spectral segments, as well as the amount of data read, cache usage, and computation required for edge computing, are all written into the graph weight expression. Nodes in the maturity-sensitive spectral segment graph are used to represent spectral segments, and the connections between nodes are used to represent the complementary or redundant relationships between spectral segments. The edge student model only receives a subset of basic spectral segments and their combined features. The full-spectrum teacher model provides the maturity probability distribution, adjacent level boundary relationships, and sample similarity relationships formed under the condition of complete spectral segments. In this embodiment, the integrity of spectral information and the edge computing load are placed in the same processing framework, so that the generation of the basic spectral segment subset has an executable data basis.

[0021] ; Where represents the joint selection score for the i-th spectral segment. This represents the contribution of the i-th spectral segment to the maturity label. This represents the average redundancy between the i-th spectral segment and the set of selected spectral segments. This represents the normalized computation cost of the i-th spectral segment in edge computation. This represents the complementarity of the i-th spectral segment to the boundaries of adjacent maturity levels. , , and These represent the weight coefficients for discriminative contribution, redundancy suppression, computational cost suppression, and boundary complementarity, respectively. Each weight coefficient is determined by the cross-validation results of the basic spectral bands selected from the training samples. For example, if a certain spectral band has a high... However, it forms a higher ratio with the selected spectral segments. At that time, its The redundancy suppression term will reduce the redundancy if the spectral band also has a high redundancy. And the calculation cost If the value is low, it can be retained as a maturity boundary compensation band.

[0022] In one embodiment, reference Figure 2The construction of the maturity-sensitive spectral map includes: calculating the discriminant contribution value between each spectral segment in the apple hyperspectral cube and the maturity level label; calculating the redundant correlation value based on spectral shape correlation and local absorption position similarity for any two spectral segments; generating a computational cost value based on the data reading volume, cache usage, and multiply-accumulate operation volume when the spectral segment participates in edge computing; and writing the discriminant contribution value, the redundant correlation value, and the computational cost value into the same spectral weight expression, so that the spectral segment selection is simultaneously constrained by maturity discrimination information, spectral segment complementarity, and edge computing load. In specific implementation, the discriminant contribution value is calculated by the separation degree of spectral segment features between samples of different maturity levels. The calculation shows that the separation degree utilizes both the inter-class mean difference and the intra-class fluctuation range to avoid the amplification of spectral contribution caused by abnormal reflection of only a few samples. The redundancy correlation value is jointly determined by the correlation of the reflection curves of the two spectral segments, the degree of overlap of absorption positions, and the degree of similarity of spectral shape differences. If the two spectral segments show the same trend of change in most samples and their absorption positions are similar, the redundancy correlation value between them is high. The calculation cost is obtained by normalizing the number of times the spectral segment data is read, the feature cache usage, and the number of times it participates in the edge student model operation. In this embodiment, whether a spectral segment enters the basic spectral segment subset is no longer determined by a single maturity correlation, but is jointly limited by maturity information, complementary information, and edge computing load.

[0023] Furthermore, the nodes in the maturity-sensitive spectral segment map record the spectral segment number, the continuous band group in which the spectral segment belongs, the maturity level contribution distribution, and the edge computing cost. The edges between nodes record redundant correlation values, complementary correlation values, and boundary collaboration identifiers. The boundary collaboration identifier is used to indicate whether two spectral segments form a complementary discriminative relationship on the adjacent level boundaries of immature and near-mature, near-mature and moderately mature, and moderately mature and overmature. When constructing the map, the edge computing processor first converts all spectral segments into standardized feature sequences, then calculates the contribution difference of each spectral segment between each level according to the maturity level label of the same batch of samples, and selects the spectral segment most sensitive to adjacent levels as candidate nodes. After the candidate nodes enter the map, they are locally clustered according to the redundant correlation values. Within the same cluster, nodes with lower computational costs and higher boundary complementarity are retained. Between different clusters, node combinations that can jointly cover the continuous maturity axis are retained. In this embodiment, the basic spectral segment subset is avoided from being concentrated in a few spectral segment regions with strong contributions but redundant information, so that the discriminative boundary between adjacent maturity levels can be preserved at the edge.

[0024] In one embodiment, reference Figure 3The knowledge constraint training of the marginal student model by the full-spectrum teacher model includes: using the maturity level probability distribution, the embedding distance of adjacent maturity level boundary samples, and the maturity similarity matrix of samples in the same batch obtained under the full-spectrum input as constraint objects; and correspondingly constraining the level probability, feature embedding, and sample similarity matrix obtained by the marginal student model under the basic spectrum subset input with the constraint objects respectively. Furthermore, an adjacent maturity level boundary preservation term is added to the training loss to ensure that the marginal student model retains the distinguishing boundaries between near-mature, moderately mature, and overmature samples under the full-spectrum condition. In specific implementation, the full-spectrum teacher model uses the complete spectrum... The segment features are used as input and output as the probability distribution of each maturity level. The marginal student model takes the basic spectral segment subset and the spectral segment difference combination features as input and outputs the corresponding probability distribution. During training, the soft probability output by the full-spectrum teacher model is used as the source of maturity boundary information, enabling the marginal student model to learn the relative judgment relationship of adjacent level samples by the full-spectrum model. The maturity similarity matrix of samples in the same batch is calculated by the distance of the samples in the intermediate feature space of the full-spectrum teacher model, which is used to constrain the marginal student model to maintain the relative order of maturity between samples. In this embodiment, the impact of information reduction caused by the basic spectral segment input on maturity boundary judgment is reduced.

[0025]

[0026] in, Let N represent the knowledge-constrained training loss, and K represent the number of training samples and the number of maturity levels. This represents the probability that the nth sample belongs to the kth maturity level in the full-spectrum teacher model. This represents the probability corresponding to the marginal student model. This represents the sample maturity similarity matrix formed by the full-spectrum teacher model. This represents the sample maturity similarity matrix formed by the marginal student model. Let B represent the sum of squares of the matrix elements, and let B represent the set of adjacent maturity level boundary sample pairs. Let m represent the embedding distance between sample a and sample b in the marginal student model, and m represent the lower bound of the boundary sample interval. , and These represent the training weights for different constraints, such as when two well-matured and over-mature boundary samples are too close together in the marginal student model. It generates boundary preservation loss and promotes feature embedding separation.

[0027] Preferably, after the full-spectrum teacher model is trained, it is not used as the main running model for edge inference. Instead, during the training phase, the maturity level probability distribution, sample similarity matrix, and boundary sample embedding distance are solidified as constraint data. The constraint data is written into the memory for use when the edge student model is trained or updated. The training samples of the edge student model are divided into regular samples and boundary samples according to the apple maturity level. Regular samples are used to maintain the basic maturity classification relationship, and boundary samples are used to strengthen the subdivision discrimination between near-mature, moderately mature, and overripe. During training, a similarity matrix is ​​constructed for the same batch of samples so that samples with similar maturity levels remain in close positions in the low-dimensional feature space of the edge student model, and samples with adjacent maturity levels but different discrimination directions are kept apart. After the model is trained, only the edge student model and the supplementary spectrum inference branch are deployed. In this embodiment, the relative maturity relationship formed under the full spectrum condition is retained while reducing the running load of the full-spectrum model at the edge.

[0028] In one embodiment, reference Figure 4 The triggering of the supplementary spectral segment inference branch includes: after the marginal student model outputs the basic inference results, calculating the confidence interval between the highest and second-highest maturity candidate levels, calculating the spectral residual after the basic spectral segment subset is reconstructed to the full-spectrum reference space, and determining the sensitive spectral segment missing risk value based on the adjacency contribution of the non-inference spectral segments in the maturity sensitive spectral segment map. When the confidence interval, the spectral residual, and the sensitive spectral segment missing risk value meet the joint triggering condition, the supplementary spectral segment associated with the spectral residual is selected to enter the secondary inference. Specifically, the confidence interval is formed by the difference between the highest and second-highest probability maturity levels in the basic inference results, the spectral residual is formed by the difference between the estimated full-spectrum curve obtained by reconstructing the basic spectral segment subset through the reference mapping and the historical full-spectrum reference curve, and the sensitive spectral segment missing risk value is calculated by the boundary contribution relationship between the non-inference spectral segments and adjacent nodes in the basic spectral segment subset. In this embodiment, the supplementary spectral segment inference is only initiated for samples with unstable basic results or insufficient spectral interpretation.

[0029] Furthermore, after completing the basic inference, the edge computing processor inputs the features corresponding to the basic spectral segment subset into the spectral shape reconstruction mapping layer. The spectral shape reconstruction mapping layer is established from the full spectrum samples during the training phase. The inputs are the reflection features, spectral shape change features, and local absorption features in the basic spectral segment subset. The output is the estimated spectral shape in the full spectrum reference space. The spectral shape residuals are formed by accumulating the differences between the estimated values ​​of each spectral segment and the mean of the full spectrum reference. If the residuals are concentrated in the neighborhood of the unparticipated spectral segments that have boundary contributions in the maturity sensitive spectral segment map, the edge computing processor marks the corresponding neighborhood as a supplementary candidate region. The confidence level is used to exclude samples whose basic results have been stable. The sensitive spectral segment missing risk value is used to exclude samples whose residuals are high but are not related to the maturity boundary. The joint triggering condition is determined by three types of quantities. In this embodiment, it is avoided that the supplementary inference is frequently triggered due to insufficient confidence or fluctuation of a single spectral shape residual.

[0030] ; in, This indicates the trigger flag for supplementary spectral segment inference branches. This indicates that the value is 1 when the condition is true and 0 when the condition is false. This represents the confidence interval between the highest and second-highest maturity candidate levels. This represents the spectral residual after the basic spectral subset is reconstructed to the full spectral reference space. This indicates the risk value for missing sensitive spectral bands. , and These represent the confidence interval boundary, spectral residual boundary, and missing risk boundary generated from the quantile positions of the training sample distribution, respectively. For example, when... Smaller Larger and When it is large, The value is set to 1 and the supplementary spectral segment inference branch is initiated when only one is available. Larger and When the corresponding boundary is not reached, It takes the value 0 and maintains the basic reasoning result.

[0031] In one embodiment, determining the basic spectral subset includes: first, selecting candidate spectral segments from the maturity-sensitive spectral segment map whose discrimination contribution values ​​meet the maturity differentiation requirements; then, performing mutual exclusion screening on the candidate spectral segments based on the redundant correlation values, and calculating complementary compensation scores for the adjacent nodes of the excluded spectral segments; if there is a maturity boundary complementary relationship between the excluded spectral segments and the retained spectral segments, then adding the adjacent spectral segments with lower computational cost values ​​to the candidate spectral segments; finally, performing combination verification on the candidate spectral segments according to the maturity level coverage integrity, so that the basic spectral subset simultaneously covers immature, near-mature, moderately mature, and overmature light. In the spectral discrimination interval, the edge computing processor sorts the spectral segments according to the joint selection score, takes the high-contribution spectral segments as the initial candidates, and then temporarily excludes the spectral segments that are highly redundant with the initial candidates. Temporary exclusion is not the same as permanent deletion. If the adjacent region of the graph corresponding to the excluded spectral segment has a complementary effect on a certain adjacent maturity level boundary, then the spectral segment with lower computational cost and lower redundancy is selected from the adjacent region to be added to the candidate set. During the combination verification, the coverage status of the candidate set on the continuous maturity axis is checked. In this embodiment, while reducing the entry of redundant spectral segments into the basic input, maturity boundary compensation spectral segments are retained.

[0032] In this embodiment, the mutual exclusion screening does not delete a fixed number of spectral segments, but rather dynamically processes them based on the redundancy correlation values ​​between spectral segment nodes and the complementary relationship of maturity boundaries. If two spectral segments have similar reflection trends across all maturity levels and the same local absorption location, only the spectral segment with lower computational cost is retained. If two spectral segments are similar on the overall spectral curve, but one of them has a local absorption difference for the moderately mature and over-mature boundary samples, the adjacent nodes of the excluded spectral segment are included in the complementary compensation score calculation. The complementary compensation score is determined by the boundary complementarity, the redundancy with the selected spectral segments, and the computational cost. The added spectral segment must form a maturity boundary coverage relationship with the selected spectral segments, and the basic spectral segment subset cannot be concentrated in the same continuous band group. After the basic spectral segment subset is determined, it is written into the memory and bound to the edge student model input layer. In this embodiment, the spectral segment screening is transformed from single-point contribution ranking to a combination selection under spectral constraints.

[0033] In one embodiment, the marginal student model includes a spectral segment combination encoding layer, a maturity boundary embedding layer, and a lightweight classification layer. The spectral segment combination encoding layer receives the basic spectral segment subset and the spectral segment difference combination features constructed from the basic spectral segment subset. The maturity boundary embedding layer constrains the interval of adjacent maturity level samples based on the boundary sample embedding distance output by the full-spectrum teacher model. The lightweight classification layer forms maturity level probabilities based on the low-dimensional features output by the maturity boundary embedding layer, so that the trained marginal student model maintains a fixed input mapping relationship with the basic spectral segment subset. Specifically, the spectral segment combination encoding layer concatenates the reflection value, adjacent spectral segment difference value, cross-spectral segment ratio, and local absorption features in the basic spectral segment subset into a low-dimensional input vector, and forms spectral segment combination features through shared weight mapping. The maturity boundary embedding layer maps the spectral segment combination features to a maturity continuous space. The lightweight classification layer converts the features in the maturity continuous space into probabilities of immature, near-mature, moderately mature, and overmature levels. In this embodiment, the model input and the basic spectral segment subset correspond stably, which facilitates inference at the edge end according to a fixed data layout.

[0034] Furthermore, when generating spectral segment combination features, the spectral segment combination encoding layer does not combine all spectral segments in pairs. Instead, it selects combination objects based on the adjacency relationships retained in the maturity-sensitive spectral segment map. If two basic spectral segments have complementary boundary markers in the map, the reflection difference, normalized ratio, and local slope difference between them are generated. If there is only a high-redundancy association between two basic spectral segments, no additional combination features are generated. The maturity boundary embedding layer receives the boundary sample embedding distance from the full-spectrum teacher model as a constraint during the training phase and outputs only low-dimensional boundary features during the inference phase. The lightweight classification layer uses the low-dimensional boundary features to form the level probability and retains the probability difference between the highest candidate level and the second highest candidate level. The probability difference is subsequently used to trigger judgment in the supplementary spectral segment. In this embodiment, the computational expansion caused by the combination of irrelevant spectral segments is reduced, and the internal features of the model directly serve the discrimination of adjacent maturity levels.

[0035] In a preferred embodiment, the selection of the supplementary spectral segments includes: based on the residual peak position of the spectral residual in the full-spectrum reference space, retrieving spectral segment nodes adjacent to the residual peak position and contributing to the maturity boundary from the maturity-sensitive spectral segment map; sorting the retrieved spectral segment nodes according to their redundancy correlation values; limiting the number of supplementary spectral segments based on computational cost; and inputting the sorted supplementary spectral segments into a supplementary spectral segment inference branch independent of the basic inference path, so that the secondary inference only uses local spectral segment information corresponding to the spectral anomaly of the current apple sample. Specifically, continuous positions in the spectral residual curve above the adaptive residual boundary are determined as residual peak regions. The edge computing processor retrieves adjacent nodes of the residual peak region in the maturity-sensitive spectral segment map. If an adjacent node has a maturity boundary contribution and is not included in the basic spectral segment subset, it enters the supplementary spectral segment candidate set. The supplementary spectral segment candidate set is sorted according to its proximity to the residual peak, boundary contribution, and computational cost. In this embodiment, the secondary inference revolves around the actual spectral differences of the current sample.

[0036] In this embodiment, the supplementary spectral segment inference branch and the basic inference path are independent of each other in terms of input, but they share the maturity level space with the basic inference result in the output layer. The supplementary spectral segment inference branch receives the reflection features of the supplementary spectral segment, the difference features between the supplementary spectral segment and the basic spectral segment, and the boundary contribution weight corresponding to the supplementary spectral segment, and generates the supplementary maturity level probability. When deduplicating the supplementary spectral segment candidate set, if multiple spectral segment nodes correspond to the same residual peak region and have high redundancy correlation values, then the spectral segment nodes with high boundary contribution weights and low computational cost values ​​are retained. If the residual peak spans multiple non-adjacent spectral segment regions, then at least one boundary contribution node is retained in each region to enter the candidate sorting. The number of supplementary spectral segments is limited by the cumulative boundary of computational cost values ​​rather than by a fixed number. In this embodiment, the secondary inference load at the edge end is controlled while ensuring the relevance of the supplementary information.

[0037] ; in, This represents the ranking score of the j-th supplementary spectral segment candidate node. This represents the contribution weight of the j-th supplementary spectral segment candidate node to the adjacent maturity level boundary. This represents the average redundancy correlation value between the candidate node of the j-th supplementary spectral segment and the selected supplementary spectral segments. This represents the normalized computational cost of the j-th supplementary spectral segment candidate node. This indicates the position of the j-th supplementary spectral segment candidate node in the spectral segment sequence. This indicates the location of the center of the current spectral residual peak. For example, a candidate node has a high... And close to When you get a higher If another candidate node has the same boundary contribution but higher redundancy with the already selected supplementary spectral band, then because The order is shifted to the next position as the number of elements increases.

[0038] In one embodiment, the consistency fusion includes: jointly encoding the maturity level probability in the basic inference result, the maturity level probability output by the supplementary spectral segment inference branch, and the boundary contribution weight of the supplementary spectral segment in the maturity-sensitive spectral segment map to form a basic evidence vector and a supplementary evidence vector; when the basic evidence vector and the supplementary evidence vector point to the same maturity level, the level probability is updated according to the boundary contribution weight; when they point to adjacent maturity levels, the embedding distance corresponding to the adjacent level boundary preservation term is called for re-discrimination and a fused maturity level is generated. Specifically, the basic evidence vector consists of the maturity level probability output by the marginal student model, the maturity boundary embedding layer output features, and the level confidence interval; the supplementary evidence vector consists of the maturity level probability output by the supplementary spectral segment inference branch, the supplementary spectral segment intermediate features, and the boundary contribution weight. If both the basic evidence vector and the supplementary evidence vector support the same maturity level, the boundary contribution weight is directly superimposed on that maturity level. If they support adjacent levels, the boundary re-discrimination process is entered. In this embodiment, the supplementary spectral segment result is avoided from simply overwriting the basic result.

[0039] Furthermore, during the consistency fusion process, the edge computing processor performs same-dimensional mapping on the basic evidence vector and the supplementary evidence vector, ensuring that both fall into the maturity boundary embedding space. Then, it calculates the distance between the two and the adjacent maturity level boundary prototypes. The maturity boundary prototypes are obtained by the full-spectrum teacher model by aggregating the boundary sample features during the training phase. If the basic evidence vector and the supplementary evidence vector point to non-adjacent levels, the non-adjacent level results are marked as an abnormal fusion state, and the basic inference results are retained as output. If they point to adjacent levels, a correction amount is generated using the boundary contribution weight and the embedding distance difference, and the correction amount is applied to the adjacent maturity level probabilities. The fused maturity level is determined by the corrected level probabilities. At the time of output, the index of the spectral residual region that triggered the supplementary inference is saved simultaneously. In this embodiment, the supplementary spectral segment information participates in the boundary correction rather than causing level jumps.

[0040] ; in, Let represent the probability that the fused result belongs to the k-th maturity level, and let represent the probability that the basic inference result belongs to the k-th maturity level. This represents the probability that the supplementary spectral segment reasoning branch belongs to the k-th maturity level. This represents the contribution weight of the supplementary spectral segment to the boundary of the k-th maturity level, where K represents the number of maturity levels. This represents a normalized summation index. For example, when both basic and supplementary reasoning support maturity levels and the maturity levels correspond... At higher levels, the fusion probability of the mature level increases after normalization. When the supplementary reasoning supports the overmature level adjacent to the basic reasoning, the fusion probability enters the boundary for further discrimination instead of directly replacing the basic level.

[0041] In one embodiment, the combined verification of maturity level coverage integrity includes: mapping the basic spectral subset onto a continuous maturity axis from immature to overmature; calculating the intra-class dispersion and adjacent level interval of samples corresponding to each maturity level under the basic spectral subset; and matching the calculation results with the adjacent level boundary embedding distance formed by the full-spectrum teacher model under full-spectrum input; when the intra-class dispersion of any maturity level exceeds its adjacent level interval, spectral nodes with complementary boundary relationships and low computational cost are backtracked from the maturity-sensitive spectral map and added to the basic spectral subset. Specifically, during the training phase, the edge computing processor projects each maturity level sample onto the feature space formed by the basic spectral subset, calculates the dispersion between samples of the same maturity level and the interval between the centers of adjacent level samples. If the fluctuation range within a certain level is greater than the interval between it and adjacent levels, it indicates that the basic spectral subset does not adequately represent the boundary of that level. The backtracking operation searches for spectral nodes with complementary contributions to that boundary in the map. In this embodiment, the basic spectral subset maintains complete coverage on the continuous maturity axis.

[0042] In this embodiment, the maturity level coverage integrity verification runs through both the determination of the basic spectral segment subset and the training of the marginal student model. After the basic spectral segment subset is initially formed, the projection results of the training samples in the basic spectral segment space are compared with the boundary embedding distance of the full-spectrum teacher model. If any boundary is compressed, such as the immature to near-mature boundary, the near-mature to moderately mature boundary, or the moderately mature to overmature boundary, then a spectral segment node that can widen the boundary and has a low computational cost is selected from the maturity-sensitive spectral segment map to be added. After the addition, the intra-class dispersion and the adjacent level interval are recalculated until each maturity level forms a relative order in the basic spectral segment space that is consistent with the full-spectrum reference boundary. If multiple candidate spectral segments meet the boundary complementarity requirement, then the spectral segment with a lower redundancy correlation value with the selected spectral segment is selected. In this embodiment, the insufficient boundary coverage of the basic spectral segment subset is corrected without returning to the full-spectrum input.

[0043] In one embodiment, the generation of the fusion maturity level includes: when the basic evidence vector and the supplementary evidence vector point to adjacent maturity levels, extracting the maturity boundary embedding layer output features of the marginal student model, projecting the intermediate features of the supplementary spectral segment inference branch onto the same embedding space, calculating the distances of the two types of features to the adjacent maturity level boundary prototypes, generating boundary correction coefficients based on the distance difference and the boundary contribution weights, and determining the fusion maturity level by redistributing the probabilities of adjacent maturity levels using the boundary correction coefficients. Specifically, the adjacent maturity level boundary prototypes are formed by aggregating the boundary sample embedding features from the full-spectrum teacher model. The basic evidence vector provides the maturity position of the current sample under the basic spectral segment subset, and the intermediate features of the supplementary spectral segment provide the local maturity position of the current sample under the residual correlation spectral segment. After the two types of features are projected onto the same embedding space, the distances to the adjacent boundary prototypes are calculated respectively. The distance difference is used to determine which side of the adjacent level the current sample is closer to. In this embodiment, the conflict between adjacent levels is resolved through the distance relationship in the boundary space.

[0044] Furthermore, the boundary correction coefficient is not directly determined by the level probability of the supplementary spectral segment inference branch, but is jointly generated by the basic embedded features, supplementary intermediate features, adjacent maturity level boundary prototypes, and supplementary spectral segment boundary contribution weights. If the basic embedded features and supplementary intermediate features are both close to the same boundary prototype, the boundary correction coefficient is assigned a higher probability to that maturity level. If the distance difference between the two features and the two adjacent boundary prototypes is small, the basic inference result is maintained and the supplementary inference identifier is retained. After determining the fusion maturity level, the edge computing processor writes the basic spectral segment subset number, supplementary spectral segment number, trigger flag, spectral residual peak position, and fusion probability into the detection record for use as sample difficulty information during subsequent batch calibration or map update. In this embodiment, the output results of the maturity boundary samples have traceable spectral segment basis and model basis.

[0045] ; in, Indicates the boundary correction coefficient. This represents the boundary prototype distance from the fused sample features to the adjacent maturity level u. This represents the boundary prototype distance from the features of the fused sample to the adjacent maturity level v. This represents the total weight of the boundary contribution corresponding to the supplementary spectral segment. This prevents positive numbers with a denominator of zero. This indicates the direction in which a sample is relatively close to rank u or rank v. If the value is positive, the probability of adjacent levels is redistributed towards level u. If the value is negative, the probability of adjacent levels is redistributed towards level v. If the result is close to 0, the hierarchical order of the basic inference results is maintained and the boundary uncertainty state is recorded.

[0046] In a preferred embodiment, the maturity-sensitive spectral map can be initially built during the offline training phase and then updated under constraints after detecting several batches of samples at the edge. The constrained update only adjusts the maturity contribution statistics and computational cost statistics of the nodes, without changing the basic spectral subset structure already bound to the input layer of the edge student model. If the new batch of samples frequently triggers the same residual peak region in multiple detection records and the supplementary spectral inference branch generates boundary corrections for the fusion maturity level multiple times, the edge computing processor marks the spectral node corresponding to the residual peak region as a candidate supplementary node. The candidate supplementary node needs to pass the maturity level coverage integrity verification in subsequent training and verification before it can enter the new basic spectral subset version. The old version of the basic spectral subset is still retained for historical detection record verification. In this embodiment, sample difficulty information can be accumulated during edge operation, while avoiding model input inconsistency caused by unconstrained changes to the map.

[0047] In a preferred embodiment, after the apple hyperspectral cube is input into the edge computing system, the data receiving interface only indexes the spectral data corresponding to the effective area of ​​the fruit surface. The edge computing processor reads the basic spectral data based on the basic spectral subset index and generates reflection features, spectral shape change features, local absorption features, and continuous spectrum envelope shift features for each spectral segment. If reflection saturation, background mixing, or continuous spectrum envelope breakage occurs in the basic spectral data, the edge computing processor marks the corresponding spectral features as low-confidence features and searches for alternative spectral segments with boundary complementarity and lower computational cost in the maturity-sensitive spectral segment map. The alternative spectral segments are only used for the correction of the basic inference input of the current sample and do not change the storage version of the basic spectral subset. The edge student model receives the corrected input vector and records the alternative spectral segment identifier simultaneously. In this embodiment, the detection process is kept continuous when local spectral segments are abnormal, and the interference of abnormal spectral segments on the maturity boundary judgment is reduced.

[0048] In a preferred embodiment, the edge computing processor establishes a hierarchical cache for the basic inference results, supplementary spectral segment inference results, and consistency fusion results of the same apple sample. The hierarchical cache includes a basic spectral segment feature cache, a boundary embedding feature cache, a supplementary spectral segment candidate cache, and a fusion probability cache. The basic spectral segment feature cache is used to avoid repeated reading of the same spectral segment in basic inference and spectral reconstruction. The boundary embedding feature cache is used to directly participate in the adjacent level re-discrimination after supplementary inference is triggered. The supplementary spectral segment candidate cache is used to record the spectral segment node sorting results corresponding to the residual peak region. The fusion probability cache is used to save the adjacent level probability allocation state before the final output. After the sample output is completed, the trigger mark, supplementary spectral segment number, and fusion maturity level are retained. The remaining intermediate features are released according to the sample end mark. In this embodiment, the edge data cache is organized around the spectral segment map and inference path to avoid the complete spectral cube occupying storage space for a long time.

[0049] In a preferred embodiment, the full-spectrum teacher model, the edge student model, and the supplementary spectral segment inference branch adopt a training and deployment separation approach. The full-spectrum teacher model learns the maturity level probability distribution, sample similarity matrix, and boundary sample embedding distance using the complete hyperspectral spectral cube at the training end. After training, only the constraint data, boundary prototype, and spectral reconstruction mapping parameters are written to the edge-end memory. The edge student model performs lightweight inference based on the basic spectral segment subset at the edge end. The supplementary spectral segment inference branch only runs when the joint triggering condition is met. If the version of the basic spectral segment subset input to the edge student model is updated, the corresponding spectral segment combination coding layer and maturity boundary embedding layer are updated synchronously. The parameters of the old version model remain bound to the old version of the basic spectral segment subset. In this embodiment, full-spectrum knowledge participates in edge detection in the form of constraint data, avoiding the edge end from running the complete full-spectrum model in each detection.

[0050] In a preferred embodiment, the edge computing system outputs a data processing status identifier simultaneously when outputting the maturity level. The data processing status identifier includes basic inference output, supplementary inference output, adjacent level fusion output, and anomaly fusion retention output. The basic inference output indicates that the basic spectral segment subset has formed a stable maturity probability. The supplementary inference output indicates that the joint triggering condition is met and the supplementary spectral segment inference branch has participated in the discrimination. The adjacent level fusion output indicates that the basic evidence vector and the supplementary evidence vector point to adjacent maturity levels and the probability is redistributed after boundary correction coefficients. The anomaly fusion retention output indicates that the basic evidence vector and the supplementary evidence vector point to non-adjacent levels and the system retains the basic inference result. The data processing status identifier, along with the maturity level, level probability, spectral residual peak position, and supplementary spectral segment number, is written into the detection record. In this embodiment, the maturity level output has a clear data processing path, which facilitates subsequent updates to the spectral statistics and analysis of the boundary sample distribution based on the detection record.

Claims

1. An edge computing system for detecting the hyperspectral maturity of apples, characterized in that, It includes a hyperspectral data receiving interface, an edge computing processor, and a memory, wherein the memory stores instructions that can be executed by the edge computing processor; The instruction causes the edge computing processor to extract spectral reflectance features, spectral shape change features, local absorption features, and continuous spectrum envelope shift features from the apple hyperspectral cubic array, and to form a maturity-sensitive spectral map with spectral segments as nodes and using maturity label contribution, spectral segment redundancy, and edge computing cost. Based on the maturity-sensitive spectral segment map, a basic spectral segment subset is determined. Based on the full-spectrum teacher model, a marginal student model that only inputs the basic spectral segment subset is trained with knowledge constraints. During marginal inference, a supplementary spectral segment inference branch is triggered based on the level confidence interval and spectral residual. The maturity level is output after the basic inference result and the supplementary spectral segment inference result are consistently fused.

2. The edge computing system for apple hyperspectral maturity detection according to claim 1, characterized in that, The construction of the maturity-sensitive spectral band map includes: calculating the discriminant contribution value between each spectral band in the apple hyperspectral cube and the maturity level label; For any two spectral bands, calculate the redundant correlation value based on spectral shape correlation and local absorption position similarity, and generate the computational cost based on the amount of data read, cache usage, and multiply-accumulate operations when the spectral bands participate in edge calculation; The discrimination contribution value, the redundancy correlation value, and the computational cost value are written into the same spectral weight expression, so that the spectral segment selection is simultaneously constrained by maturity discrimination information, spectral segment complementarity relationship, and edge computing load.

3. The edge computing system for apple hyperspectral maturity detection according to claim 1, characterized in that, The knowledge constraint training of the full-spectrum teacher model on the marginal student model includes: using the maturity level probability distribution obtained under full-spectrum input, the embedding distance of adjacent maturity level boundary samples, and the maturity similarity matrix of samples in the same batch as constraint objects. The rank probability, feature embedding, and sample similarity matrix obtained by the marginal student model under the input of the basic spectral segment subset are respectively constrained with the constraint object, and an adjacent maturity rank boundary preservation term is added to the training loss; The marginal student model retains the distinguishing boundaries between near-mature, moderately mature, and over-mature samples under full-spectrum conditions.

4. The edge computing system for apple hyperspectral maturity detection according to claim 1, characterized in that, The triggering of the supplementary spectral segment inference branch includes: after the marginal student model outputs the basic inference result, calculating the level confidence interval between the highest candidate level of maturity and the second highest candidate level; Calculate the spectral shape residual after the basic spectral segment subset is reconstructed to the full spectral reference space, and determine the sensitive spectral segment missing risk value based on the adjacency contribution of the spectral segments that did not participate in the inference in the maturity sensitive spectral segment map; When the confidence interval, the spectral residual, and the risk value of missing sensitive spectral segments meet the joint triggering condition, a supplementary spectral segment associated with the spectral residual is selected for secondary inference.

5. The edge computing system for apple hyperspectral maturity detection according to claim 2, characterized in that, The determination of the basic spectral segment subset includes: firstly, selecting candidate spectral segments from the maturity-sensitive spectral segment map whose discrimination contribution values ​​meet the maturity differentiation requirements; Then, mutual exclusion screening is performed on the candidate spectral segments based on the redundant correlation values, and complementary compensation scores are calculated for the adjacent nodes of the excluded spectral segments. If there is a maturity boundary complementary relationship between the excluded spectral segments and the retained spectral segments, the adjacent spectral segments with low computational cost values ​​are added to the candidate spectral segments. Finally, the candidate spectral segments are combined and verified according to the maturity level coverage integrity, so that the basic spectral segment subset simultaneously covers the spectral discrimination intervals of immature, near-mature, moderately mature, and overmature.

6. The edge computing system for apple hyperspectral maturity detection according to claim 3, characterized in that, The marginal student model includes a spectral combination coding layer, a maturity boundary embedding layer, and a lightweight classification layer; The spectral band combination coding layer receives the basic spectral band subset and the spectral band differential combination features constructed from the basic spectral band subset; The maturity boundary embedding layer constrains the interval between adjacent maturity level samples based on the boundary sample embedding distance output by the full-spectrum teacher model. The lightweight classification layer forms maturity level probabilities based on the low-dimensional features output by the maturity boundary embedding layer, so that the trained marginal student model maintains a fixed input mapping relationship with the basic spectral segment subset.

7. The edge computing system for apple hyperspectral maturity detection according to claim 4, characterized in that, The selection of the supplementary spectral bands includes: based on the position of the residual peaks in the full-spectrum reference space of the spectral residuals; From the maturity-sensitive spectral segment map, retrieve spectral segment nodes that are adjacent to the position of the residual peak and have a maturity boundary contribution, sort the retrieved spectral segment nodes according to the redundancy correlation value, and then supplement the number of spectral segments according to the calculation cost limit. The sorted supplementary spectral segments are input into the supplementary spectral segment inference branch, which is independent of the basic inference path, so that the secondary inference only uses the local spectral segment information corresponding to the spectral anomaly of the current apple sample.

8. The edge computing system for apple hyperspectral maturity detection according to claim 4, characterized in that, The consistency fusion includes: jointly encoding the maturity level probability in the basic inference result, the maturity level probability output by the supplementary spectral segment inference branch, and the boundary contribution weight of the supplementary spectral segment in the maturity sensitive spectral segment map to form a basic evidence vector and a supplementary evidence vector. When the basic evidence vector and the supplementary evidence vector point to the same maturity level, the level probability is updated according to the boundary contribution weight. When both point to adjacent maturity levels, the embedding distance corresponding to the adjacent level boundary preservation item is called for re-discrimination and a fusion maturity level is generated.

9. The edge computing system for apple hyperspectral maturity detection according to claim 5, characterized in that, The combined verification of the maturity level coverage integrity includes: mapping the basic spectrum subset to a continuous maturity axis from immature to overmature, calculating the intraclass dispersion and adjacent level interval of the samples corresponding to each maturity level under the basic spectrum subset, and matching the calculation results with the adjacent level boundary embedding distance formed by the full spectrum teacher model under the full spectrum input; When the intraclass dispersion of any maturity level exceeds the interval between its adjacent levels, a spectral node with a boundary complementary relationship and low computational cost is backtracked from the maturity-sensitive spectral segment map and added to the basic spectral segment subset.

10. The edge computing system for apple hyperspectral maturity detection according to claim 8, characterized in that, The generation of the fusion maturity level includes: when the basic evidence vector and the supplementary evidence vector point to adjacent maturity levels, extracting the maturity boundary embedding layer output features of the marginal student model, projecting the intermediate features of the supplementary spectral inference branch to the same embedding space, and calculating the distances of the two types of features to the adjacent maturity level boundary prototypes respectively. A boundary correction coefficient is generated based on the distance difference and the boundary contribution weight, and the fusion maturity level is determined by redistributing the probabilities of adjacent maturity levels using the boundary correction coefficient.

Images