Oil tea fruit maturity online discrimination method based on deep learning edge calculation

Abstract

The invention discloses an oil tea fruit maturity online discrimination method based on deep learning edge calculation, and the method comprises the steps: synchronously collecting the original spectrum, space image and growth environment associated data of oil tea fruits through an edge end platform, and dividing the data into corresponding feature subsets through a data distribution unit; a deep residual spectrum analysis model is called to mine spectrum deep maturity associated features, an adaptive weight classification algorithm is adopted to screen image features and distribute weights to construct a multi-dimensional vector, and multi-scale feature fusion is realized through a feature pyramid recognition network to generate a discrimination feature map; and finally, outputting a maturity category result at an edge end and storing related parameters. Wherein the deep residual model reinforces feature extraction through residual block jump connection, the adaptive weight algorithm dynamically optimizes feature weights, the feature pyramid network reinforces key information through cross-layer interaction, feature processing and fusion are completed in each step in stages, and efficient and accurate online judgment of the oil tea fruit maturity in a field scene is achieved.

Description

Technical Field

[0001] This invention relates to the field of Camellia oleifera fruit maturity determination technology, and in particular to an online method for determining the maturity of Camellia oleifera fruit based on deep learning edge computing. Background Technology

[0002] Camellia oleifera, an important woody oilseed crop in my country, relies heavily on fruit maturity to determine the yield and quality of its oil. Accurate and efficient maturity assessment is crucial for optimizing harvesting timing and improving industry efficiency. Traditional methods for determining the maturity of Camellia oleifera fruit depend on manual observation of appearance features such as peel color and shape. This is heavily influenced by subjective experience, resulting in low accuracy and efficiency, making it unsuitable for large-scale planting. With the development of intelligent agriculture, spectral analysis and image recognition technologies are increasingly being applied to crop maturity detection. However, existing technologies largely rely on cloud computing for data processing, leading to issues such as data transmission delays, privacy risks, and network dependence in edge scenarios, failing to meet the real-time assessment needs in the field. Deep learning-based feature extraction and classification algorithms have demonstrated excellent performance in crop detection. Combining these with the local data processing capabilities of edge computing, an online assessment method integrating data acquisition, feature analysis, and intelligent assessment has emerged as an important direction for solving the problem of efficient Camellia oleifera fruit maturity detection.

[0003] Existing technologies for determining the maturity of camellia fruit have two significant drawbacks: First, the accuracy of feature extraction and fusion is insufficient. Traditional spectral analysis models struggle to uncover deep features strongly correlated with maturity in the data. Image feature processing often employs fixed weight allocation, failing to dynamically adjust feature importance based on the growth status of the camellia fruit. This results in a one-sided judgment criterion after multi-source feature fusion, making it difficult to adapt to the feature differences of different growth environments and varieties of camellia fruit. Second, the system deployment and real-time adaptability are poor. Existing methods do not fully integrate the characteristics of edge hardware resources for model optimization. Either the complexity of the model leads to stuttering and excessively high judgment latency at the edge, or the model structure is simplified to adapt to hardware, sacrificing judgment performance. Furthermore, there is a lack of a dedicated integrated edge computing platform for camellia fruit maturity determination, resulting in poor coordination between data acquisition, processing, and decision-making, making it impossible to achieve efficient online determination in field scenarios. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides an online method for determining the maturity of camellia fruit based on deep learning edge computing.

[0005] The technical solution adopted in this invention is an online method for determining the maturity of camellia fruit based on deep learning edge computing, comprising the following steps: S1, acquiring the original spectral data and spatial image data of the target area of ​​camellia fruit through the multispectral acquisition module of the edge-end intelligent camellia identification and analysis platform, and simultaneously acquiring the related parameters of the camellia fruit growth environment; S2, transmitting the acquired original spectral data and spatial image data to the built-in data allocation unit at the edge end, and dividing the spectral feature data subset and the image feature data subset according to preset rules; S3, calling the deep residual camellia spectral analysis model to perform layered analysis on the spectral feature data subset. The system employs a multi-level feature extraction mechanism, using residual block jump connections to mine deep maturity-related features in spectral data. In step S4, an adaptive weighted camellia maturity classification algorithm is used to filter and weight subsets of image feature data, constructing a multi-dimensional image feature vector. In step S5, the deep spectral features and weighted image feature vectors are input into a feature pyramid camellia maturity recognition network, generating a maturity discrimination feature map through multi-scale feature fusion and pyramid hierarchical mapping. Finally, the edge-end intelligent camellia identification and analysis platform outputs the camellia fruit maturity category results based on the discrimination feature map, while simultaneously storing feature data and discrimination process parameters.

[0006] Furthermore, the expression for the deep residual Camellia oleifera spectral analysis model is as follows: ,in, To output eigenvalues ​​for spectral analysis, The original spectral wavelength parameters of camellia fruit, Optimize the parameter set for model training. For activation function, For the first Feature weight coefficients of each residual block For the first Each residual calculation module For environmental parameter correction factors, It is a spectrum-environment correlation mapping function. This is a set of parameters related to the growth environment of camellia oleifera fruit. This represents the number of residual blocks.

[0007] Furthermore, the expression for the adaptive weighted camellia oleifera maturity classification algorithm is as follows: ,in, The weighted image feature vector. The original image feature vector, This is the weight adaptive adjustment factor. This is the function for calculating feature similarity. This is the standard feature vector of maturity. The number of image feature dimensions. These are the feature gradient correction coefficients. This represents the gradient value of the eigenvector.

[0008] Furthermore, the expression for the feature pyramid camellia oleifera maturity recognition network is as follows: ,in, For the first pyramid Layer fusion feature map, For deep spectral characteristics, For weighted image feature vectors, For convolution operations, These are the convolution kernel parameters for spectral features and image features, respectively. For upsampling operation, For feature element-wise multiplication operation For the first Layer feature enhancement matrix.

[0009] Furthermore, the parameter optimization expression of the edge-end camellia oil intelligent identification and analysis platform is as follows: ,in, For the optimal combination of platform operating parameters, For platform parameter set, For maturity standard datasets, For parameter balance coefficients, For parameters The corresponding discrimination delay value, For parameters The deviation value between the discrimination result and the standard dataset.

[0010] Furthermore, the comprehensive decision expression for online determination of the maturity of camellia fruit is as follows: ,in, For maturity category results, For class determination sign function, For the first The decision weights of the features in the pyramid layer. For the first The discriminative contribution value of the layer pyramid features. The number of characteristic layers of the pyramid. For process parameter correction coefficients, These are parameters used to determine the stability of the process.

[0011] Further, S3 includes the following sub-steps: S31, starting the input layer of the deep residual camellia oleifera spectral analysis model, dividing the spectral feature data subset into multiple continuous feature segments according to wavelength intervals, each feature segment corresponding to the initial input channel of the model; S32, performing preliminary extraction of features from each channel through the convolutional layer in the model to generate a basic spectral feature map, which is then transmitted to the first residual block for feature enhancement, and retaining the original feature information through residual connections; S33, sequentially performing hierarchical iterative processing on the enhanced features through subsequent residual blocks, adjusting the feature distribution through batch normalization operations in each residual block, and enhancing the nonlinear expression capability of features through activation functions; S34, performing global average pooling operations on the features output by the last residual block to obtain deep maturity association features with uniform dimensions, and transmitting them to the feature fusion unit.

[0012] Further, step S4 includes the following sub-steps: S41, decompose the image feature data subset into feature dimensions to obtain multiple independent feature components such as color features, texture features, and morphological features, each feature component corresponding to a different correlation dimension of the maturity of camellia fruit; S42, calculate the correlation value between each feature component and the maturity category through the feature importance evaluation module in the adaptive weighted camellia maturity classification algorithm, and sort and select core feature components based on the correlation value; S43, dynamically allocate the weight coefficient of each core feature component according to the selection result, assigning a higher weight coefficient to the feature component with a higher correlation value, while setting a weight adjustment threshold to limit the occurrence of extreme weights; S44, reorganize the weighted core feature components according to the preset dimensional order to construct a multi-dimensional image feature vector that can comprehensively reflect the maturity information of camellia fruit.

[0013] Further, S5 includes the following sub-steps: S51, inputting deep spectral features and multi-dimensional image feature vectors into different input branches of the feature pyramid camellia oleifera maturity recognition network, and performing feature dimension matching processing through branch convolutional layers to ensure that the two types of features have the same dimensional specifications; S52, activating the network's multi-scale feature fusion module, performing downsampling and upsampling operations on the matched features at different scales to generate a set of multi-scale feature maps, with each scale feature map corresponding to maturity feature information at different levels; S53, stacking the multi-scale feature maps in hierarchical order through the pyramid hierarchical mapping unit, and performing information interaction and complementarity of features at different levels through a cross-layer connection mechanism to strengthen the expression of maturity feature labeling; S54, performing final convolutional compression processing on the stacked feature maps to generate a discriminative feature map that can be directly used for maturity discrimination, and transmitting it to the decision output unit at the edge.

[0014] A method for online maturity assessment of camellia oleifera fruit based on deep learning edge computing is implemented through different units, including: a multi-source data collaborative acquisition unit, used to simultaneously acquire raw spectral data, spatial image data, and growth environment-related parameters of camellia oleifera fruit, and establish bidirectional data transmission connections with the spectral acquisition module, image acquisition module, and environmental sensors through a data interface; a feature data partitioning and processing unit, which receives the data output from the multi-source data collaborative acquisition unit, divides the spectral feature data subset and the image feature data subset according to data type, and transmits them to the corresponding feature processing units respectively; and a deep feature extraction and weighting unit, which incorporates a deep residual camellia oleifera spectral analysis model and an adaptive weighted camellia oleifera maturity classification algorithm. The system receives feature data after partitioning, performs deep spectral feature extraction and image feature weighting, and outputs two types of feature vectors. A multi-scale feature fusion and recognition unit, equipped with a feature pyramid camellia maturity recognition network, receives feature vectors output by the deep feature extraction and weighting unit, and generates a discriminative feature map through multi-scale fusion and hierarchical mapping. A maturity decision output unit performs maturity category determination based on the discriminative feature map, outputs the determination result through a display interface, and establishes a connection with the data storage unit. A full-process parameter storage unit receives raw data, intermediate feature data, and operating parameters transmitted from each functional unit, stores them according to timestamps and data types, and supports subsequent data retrieval and traceability.

[0015] Beneficial Effects: This invention proposes an online maturity determination method for camellia oleifera fruit based on deep learning edge computing. This method collaboratively collects spectral, image, and growth environment data of camellia oleifera fruit through an integrated edge platform. After partitioning and processing, it leverages the hierarchical extraction and residual connection mechanism of a deep residual model to mine deep maturity-related features in the spectrum. An adaptive weighting algorithm dynamically selects core image features and assigns weights, and then a feature pyramid network is used to achieve multi-scale feature fusion to generate a discriminative feature map. Finally, maturity determination is completed at the edge, and all process parameters are stored, effectively overcoming the shortcomings of existing technologies. This invention compensates for the limitations of fixed weights and shallow features, improving the comprehensiveness and accuracy of the discrimination criteria. Addressing the issue of poor system deployment and real-time adaptability, it relies on a local data processing architecture at the edge, achieving integrated data collection, processing, and decision-making through efficient connection of various functional units, avoiding cloud transmission delays and network dependencies. Simultaneously, platform parameter optimization balances operational efficiency and discrimination performance, meeting the online discrimination needs of field scenarios. The storage of all process parameters also provides support for subsequent model iteration and data traceability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0017] Figure 2 This is a flowchart of method step S3 of the present invention;

[0018] Figure 3 This is a flowchart of method step S4 of the present invention;

[0019] Figure 4 This is a flowchart of step S5 of the method of the present invention;

[0020] Figure 5 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, the online method for determining the maturity of camellia fruit based on deep learning edge computing includes the following steps:

[0023] S1, through the multispectral acquisition module of the edge-end intelligent identification and analysis platform for camellia oleifera, acquires the original spectral data and spatial image data of the target area of ​​camellia oleifera fruit, and simultaneously collects the parameters related to the growth environment of camellia oleifera fruit;

[0024] Specifically, step S1 involves using the multispectral acquisition module on the edge-end intelligent identification and analysis platform for camellia oleifera fruit to collect data from the target area. The spectral detection range of the acquisition device is set to 400 to 1000 nanometers, including key bands of visible and near-infrared light. The image acquisition resolution is configured to 1920×1080 pixels, and the frame rate is maintained at 30 frames per second to ensure the complete capture of the details and spectral response of the camellia oleifera fruit surface. During the acquisition process, the multispectral acquisition module uses a gimbal control to scan the camellia oleifera fruit plant. Each acquisition covers a detection area with a radius of 5 to 8 meters. Simultaneously, the environmental parameter acquisition module is triggered to obtain three core growth environment parameters: air humidity, light intensity, and temperature. The humidity acquisition accuracy is ±2%RH, the light intensity acquisition range is 0 to 100,000 lux, and the temperature acquisition accuracy is ±0.5℃. During data collection, laser positioning technology is used to locate the target area of ​​the camellia fruit, eliminating background interference from leaves, branches, etc. For each target area, 10 to 15 sets of raw spectral data and spatial image data are continuously collected and temporarily stored through a data caching module to ensure the continuity and integrity of data transmission. This step provides multi-dimensional and high-precision raw data support for subsequent feature extraction and discriminant analysis. Its implementation quality directly affects the accuracy of subsequent maturity determination. By synchronously collecting environmental parameters, necessary environmental correlation information is provided for subsequent model correction.

[0025] S2 transmits the acquired raw spectral data and spatial image data to the built-in data allocation unit at the edge, and divides the spectral feature data subset and the image feature data subset according to preset rules;

[0026] Specifically, step S2 involves transmitting the raw spectral data and spatial image data collected and cached in step S1 to the data distribution unit via the high-speed data transmission interface built into the edge. The processor of this unit is configured with a main frequency of 2.8 GHz and a data processing bandwidth of no less than 10 Gbps to ensure the rapid reception and distribution of massive amounts of data. The data allocation unit divides the data according to preset classification rules. Spectral data is divided into six continuous bands of spectral feature data subsets based on wavelength range: 400-500 nm, 501-600 nm, 601-700 nm, 701-800 nm, 801-900 nm, and 901-1000 nm. Each subset includes key parameters such as spectral intensity, peak wavelength, and full width at half maximum (FWHM) for the corresponding band. Spatial image data is divided into three subsets based on feature type: color feature data subset, texture feature data subset, and morphological feature data subset. The color feature data subset includes parameters such as pixel value distribution, hue mean, and saturation variance for the RGB three-color channels. The texture feature data subset includes parameters such as gray-level co-occurrence matrix, texture entropy, and contrast. The morphological feature data subset includes parameters such as area, perimeter, roundness, and contour complexity of the camellia fruit. During the partitioning process, the data allocation unit performs integrity checks on each data subset using a verification algorithm. The verification pass rate must reach 100% to ensure that the partitioned data is free of missing or incorrect data. After partitioning, the data subsets are stored in dedicated data partitions at the edge, with each partition configured with a storage capacity of 128GB. RAID5 array storage is used to ensure data security. This step achieves orderly classification and efficient management of the original data, providing a clear data foundation for subsequent targeted feature processing.

[0027] S3, calls the deep residual camellia oleifera spectral analysis model to perform hierarchical feature extraction on the subset of spectral feature data, and mines the deep maturity correlation features in the spectral data through the residual block jump connection mechanism;

[0028] Specifically, step S3 involves calling the pre-deployed deep residual camellia oleifera spectral analysis model at the edge. This model comprises 12 residual blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and an activation function layer. The model has 6 input channels, corresponding to the six spectral band data subsets divided in step S2, and 256 output channels, used to output deep maturity correlation features. During model execution, each spectral feature data subset is first input into its corresponding initial input channel. The data dimension is compressed to 128 dimensions through the first convolutional layer, and after batch normalization, it is input into the activation function layer for nonlinear transformation, and then transmitted to the first residual block. Each residual block further extracts and transforms features through convolutional layers, and simultaneously, through jump connections, the input features are directly superimposed onto the output of the residual block, effectively mitigating the gradient vanishing problem and preserving the original feature information. The first six residual blocks are used to extract low- and mid-order features from the spectral data, while the latter six residual blocks focus on mining high-order, deep maturity-related features. The convolutional stride of each residual block is dynamically adjusted to 1 or 2 according to the feature extraction requirements. When the stride is 2, zero-padding is used to maintain the consistency of the feature map size. The model is trained for 1000 iterations, with an initial learning rate of 0.001, which decays to 0.1 times the original rate every 200 iterations. The regularization coefficient is set to 0.0001 to avoid overfitting. Finally, a global average pooling layer is used to convert the feature maps output from the 12 residual blocks into 256-dimensional one-dimensional feature vectors, i.e., deep maturity-related features. This step, through hierarchical feature extraction and residual connection mechanisms, fully mines the deep information related to the maturity of camellia fruit in the spectral data, removes irrelevant and interfering features, and provides high-value spectral feature support for subsequent feature fusion.

[0029] S4. An adaptive weighted Camellia oleifera maturity classification algorithm is used to filter features and assign weights to a subset of image feature data to construct a multi-dimensional image feature vector.

[0030] Specifically, step S4 involves processing the image feature data subset divided in step S2 using an adaptive weighted camellia maturity classification algorithm. This algorithm comprises three core modules: a feature importance assessment module, a dynamic weight allocation module, and a feature vector reorganization module. First, the feature importance assessment module decomposes the three image feature data subsets (color, texture, and morphology) into 28 independent feature components: 8 color components, 12 texture components, and 8 morphology components. The correlation between each feature component and the camellia fruit maturity category is calculated using the Pearson correlation coefficient, with a confidence level of 95%. Fifteen core feature components with a correlation value greater than 0.6 are selected. Subsequently, the dynamic weight allocation module assigns weights based on the correlation values ​​of the selected core feature components. After normalization, the weight values ​​fall within the range of 0.05 to 0.25. The feature component with the highest correlation value is assigned the highest weight of 0.25. Simultaneously, a weight adjustment threshold of 0.05 is set to limit the minimum weight to no less than 0.05, preventing individual feature components from being excessively ignored. Finally, the feature vector recombination module recombines the 15 weighted core feature components into a 15-dimensional multi-dimensional image feature vector in the order of color, texture, and morphology. During the recombination process, feature standardization is used to unify the numerical range of each feature component to the [0,1] interval, ensuring the fusion of different types of features. The algorithm runs at 500 iterations with a convergence threshold of 0.0001. Weight adjustment stops when the weight change is less than the convergence threshold for 10 consecutive steps. This step, through precise feature selection, dynamic weight allocation, and scientific vector recombination, constructs an image feature vector that accurately reflects the maturity information of camellia fruit, thus improving the contribution of image features to maturity determination.

[0031] S5 inputs deep spectral features and weighted image feature vectors into the feature pyramid Camellia oleifera maturity recognition network, and generates maturity discrimination feature maps through multi-scale feature fusion and pyramid hierarchical mapping.

[0032] Specifically, step S5 involves simultaneously inputting the 256-dimensional deep spectral features obtained in step S3 and the 15-dimensional multi-dimensional image feature vector constructed in step S4 into the feature pyramid camellia oleifera maturity recognition network. This network comprises four core layers: an input branch layer, a multi-scale feature fusion layer, a pyramid level mapping layer, and an output layer. The overall network depth is set to 8 layers, and the number of parameters is controlled within 5 million to adapt to edge hardware resources. First, the input branch layer includes two independent branches, which process the deep spectral features and the multi-dimensional image feature vector respectively. A 1×1 convolutional layer is used to compress the dimension of the deep spectral features from 256 dimensions to 64 dimensions, and expand the dimension of the multi-dimensional image feature vector from 15 dimensions to 64 dimensions, achieving dimensional matching between the two types of features. Subsequently, the multi-scale feature fusion layer downsamples the matched features using three convolutional kernels of different scales (3×3, 5×5, and 7×7), generating feature maps of three scales (64×64, 32×32, and 16×16). Then, an upsampling operation restores the smaller-scale feature maps to 64×64 size, which are then element-wise added to the original scale feature maps to generate a multi-scale fused feature map set. Next, the pyramid hierarchical mapping layer divides the fused feature maps into four levels according to priority. Each level corresponds to maturity feature information of different importance. Information interaction between feature maps of different levels is achieved through a cross-layer connection mechanism. Each feature map undergoes two 3×3 convolutional layers to enhance feature representation, while an attention mechanism highlights key feature regions. Finally, the output layer compresses and fuses the feature maps of the four levels into a 64×64 maturity discrimination feature map through a 1×1 convolutional layer. Each pixel in the discrimination feature map corresponds to the maturity feature intensity value of the target region of the camellia fruit. This step fully integrates the advantages of spectral features and image features through multi-scale feature fusion, hierarchical mapping and cross-layer interaction, strengthens the expression of key maturity features, and provides high-precision feature support for the final maturity discrimination.

[0033] The S6 edge-end intelligent identification and analysis platform for camellia oleifera outputs the maturity category results of camellia oleifera fruit based on the discrimination feature map, and stores the feature data and discrimination process parameters.

[0034] Specifically, in step S6, the edge-end intelligent camellia oil identification and analysis platform receives the 64×64 maturity discrimination feature map generated in step S5. The built-in decision module analyzes and processes the feature map. The decision module employs a classification strategy based on feature intensity thresholds. It pre-determines the feature intensity threshold ranges corresponding to the three maturity levels (immature, mature, and overmature) using the training dataset, which are [0, 0.3), [0.3, 0.7), and [0.7, 1], respectively. The decision module first calculates the mean feature intensity of all pixels in the discrimination feature map. If the mean is in the range [0, 0.3), it is judged as immature; in the range [0.3, 0.7), it is judged as mature; and in the range [0.7, 1], it is judged as overmature. Simultaneously, it calculates the standard deviation of the feature intensity. If the standard deviation is less than 0.1, the discrimination result is considered stable. If the standard deviation is greater than or equal to 0.1, the feature pyramid network is called again for secondary feature extraction and discrimination to ensure the reliability of the discrimination result. The discrimination results are output in real time via the platform's display interface in a combination of text and icons, with an output latency of no more than 500 milliseconds, meeting the requirements for online discrimination. Simultaneously, the platform's storage module categorizes and stores raw spectral data, spatial image data, deep spectral features, multi-dimensional image feature vectors, discrimination feature maps, maturity category results, and operational parameters for each step, based on timestamps (accurate to milliseconds) and the target region number of the camellia fruit. The storage format adopts standard binary format, with the amount of data stored for each target region controlled within 5MB. The total capacity of the storage module is configured to 1TB, supporting data cyclic overwriting and manual export functions. This step achieves rapid output of maturity results and complete storage of the entire process data, meeting the needs of real-time field discrimination and providing rich raw data support for subsequent model optimization, data traceability, and method improvement.

[0035] Preferably, the expression for the deep residual camellia oleifera spectral analysis model is: ,in, To output eigenvalues ​​for spectral analysis, The original spectral wavelength parameters of camellia fruit, Optimize the parameter set for model training. For activation function, For the first Feature weight coefficients of each residual block For the first Each residual calculation module For environmental parameter correction factors, It is a spectrum-environment correlation mapping function. This is a set of parameters related to the growth environment of camellia oleifera fruit. This represents the number of residual blocks.

[0036] Specifically, the deep residual Camellia oleifera spectral analysis model first calls the original spectral wavelength parameters of the Camellia oleifera fruit during runtime. These parameters cover the key detection band from 400 to 1000 nanometers, including the spectral intensity data corresponding to each wavelength point. The model training optimization parameter set includes key parameters such as learning rate, regularization coefficient, and number of iterations. The learning rate is initially set to 0.001, decreasing to 0.1 times its original value every 200 iterations. The regularization coefficient is set to 0.0001, and the number of iterations is fixed at 1000. These parameters are used to control the convergence speed and fitting effect during model training. The ReLU function is used to enhance the nonlinear expressive power of features and avoid gradient saturation during model training. The feature weight coefficients of the i-th residual block are dynamically allocated according to the residual block hierarchy. The weight coefficients of the first 6 residual blocks range from 0.05 to 0.15, and the weight coefficients of the last 6 residual blocks range from 0.1 to 0.2, ensuring that deep features are given sufficient attention. Each residual calculation module consists of two 3×3 convolutional layers, a batch normalization layer, and an activation function layer. It extracts features and transforms dimensions through convolution operations. The environmental parameter correction coefficient is set to 0.03 to quantify the influence of growth environment-related parameters on spectral features. The spectral-environment correlation mapping function establishes the correlation between environmental factors and spectral data by integrating three environmental parameters: air humidity, light intensity, and temperature. The set of growth environment-related parameters includes real-time data collected from the above three parameters. The number of residual blocks is fixed at 12. Through hierarchical residual calculation and environmental parameter correction, the deep maturity correlation features in the spectral data are accurately extracted, improving the model's adaptability to camellia fruit spectral data under different environmental conditions.

[0037] Preferably, the expression for the adaptive weighted camellia oleifera maturity classification algorithm is: ,in, The weighted image feature vector. The original image feature vector, This is the weight adaptive adjustment factor. This is the function for calculating feature similarity. This is the standard feature vector of maturity. The number of image feature dimensions. These are the feature gradient correction coefficients. This represents the gradient value of the eigenvector.

[0038] Specifically, the adaptive weighted camellia oleifera maturity classification algorithm processes original image feature vectors comprising 28 feature components across three categories: color, texture, and morphology. Each feature component is standardized, with values ​​uniformly ranging from [0,1]. An adaptive weight adjustment factor of 1.2 is used to regulate the influence of feature similarity on weight allocation, ensuring the sensitivity of weight adjustment remains within a reasonable range. The feature similarity calculation function employs the Pearson correlation coefficient algorithm with a confidence level of 95%. By calculating the correlation between each original image feature vector and the maturity standard feature vector, the correlation strength between feature components and maturity categories is quantified. The maturity standard feature vector is trained based on 10,000 sets of camellia oleifera fruit image samples labeled with maturity levels, including typical features for immature, mature, and over-ripe levels. The image feature dimension is 28, corresponding to 28 independent image feature components. A feature gradient correction coefficient of 0.08 is used to correct feature vector gradient values, avoiding interference from gradient anomalies in weight allocation. The feature vector gradient value is obtained by calculating the rate of change of the feature component in the sample space, reflecting the discriminative ability of the feature component. During algorithm execution, 15 core feature components with a correlation value greater than 0.6 are first selected by feature similarity calculation. Then, normalized weights are assigned based on the correlation value, with the weight values ​​ranging from 0.05 to 0.25. At the same time, the weight vector is optimized by feature gradient correction. Finally, a weighted image feature vector is generated. This vector can highlight the contribution of core features and weaken the interference of irrelevant features, providing high-quality image feature support for subsequent feature fusion.

[0039] Preferably, the expression for the feature pyramid camellia oleifera maturity recognition network is: ,in, For the first pyramid Layer fusion feature map, For deep spectral characteristics, For weighted image feature vectors, For convolution operations, These are the convolution kernel parameters for spectral features and image features, respectively. For upsampling operation, For feature element-wise multiplication operation For the first Layer feature enhancement matrix.

[0040] Specifically, the fusion mechanism and operating parameters of the feature pyramid camellia oil maturity recognition network are as follows: the deep spectral features processed by the network are 256-dimensional vectors, and the weighted image feature vectors are 15-dimensional vectors. Both are unified in dimension through convolution operations in the input branches, transforming them into 64-dimensional feature vectors. The convolution operation uses three different kernel parameters: 3×3 for spectral features and 1×1 for image features. Convolution operations achieve feature dimension transformation and preliminary feature extraction. The upsampling operation uses bilinear interpolation to enlarge small-scale feature maps by a factor of 2, ensuring size consistency across different scales and facilitating feature fusion calculations. Element-wise feature multiplication is used to enhance the response intensity of key feature regions. By performing element-wise operations between the fused feature map and the l-th layer feature enhancement matrix, the expressive power of maturity-related features is improved. The l-th layer feature enhancement matrix is ​​constructed based on the training dataset and includes four levels. The dimension of the feature enhancement matrix at each level corresponds to the dimension of the feature at that level. Figure 1 The matrix elements range from 0.8 to 1.2, used to dynamically adjust the weights of features at different levels. During network operation, the spectral features and image features after dimension matching are first processed by multi-scale convolution to generate feature maps of three scales: 64×64, 32×32, and 16×16. Then, multi-scale fusion is achieved by upsampling and feature addition. Subsequently, information exchange between the four levels of feature maps is achieved through a cross-layer connection mechanism. Finally, the fusion feature map of the l-th layer of the pyramid is generated by element-wise multiplication and convolution compression. This feature map integrates maturity features from multiple sources and at multiple scales, providing strong feature support for accurate discrimination.

[0041] Preferably, the parameter optimization expression of the edge-end camellia oil intelligent identification and analysis platform is as follows: ,in, For the optimal combination of platform operating parameters, For platform parameter set, For maturity standard datasets, For parameter balance coefficients, For parameters The corresponding discrimination delay value, For parameters The deviation value between the discrimination result and the standard dataset.

[0042] Specifically, the edge-end intelligent camellia identification and analysis platform's parameter set includes 12 key parameters, such as data acquisition frequency, processor operating frequency, memory allocation size, and model inference accuracy. Each parameter has a set reasonable value range, with the data acquisition frequency ranging from 10 to 30 frames per second, the processor operating frequency ranging from 2.0 to 3.0 GHz, the memory allocation size ranging from 4 to 8 GB, and the model inference accuracy supporting both half-precision and single-precision modes. The maturity standard dataset includes 5,000 sets of camellia fruit samples from different growing environments, varieties, and maturity levels. Each set of samples includes complete raw data, feature data, and maturity labeling results, providing a reliable reference for parameter optimization. The parameter balance coefficient is set to 0.4 to balance the weights of discrimination delay and discrimination bias, enabling the platform to achieve an optimal balance between operating efficiency and discrimination accuracy. The discrimination latency value is obtained by monitoring the total time from data input to result output on the platform in real time, in milliseconds, with a target of less than 500 milliseconds. The deviation between the discrimination result and the standard dataset is obtained by calculating the consistency between the platform output and the standard labeled result, quantified using an accuracy index, with a target value of no less than 95%. The parameter optimization process uses a grid search algorithm to traverse all possible combinations of the platform parameter set, calculate the weighted objective function value corresponding to each parameter combination. The weighted objective function is the product of the parameter balance coefficient and the discrimination latency value plus the product of (1 - parameter balance coefficient) and the discrimination deviation value. Finally, the parameter combination with the smallest weighted objective function value is selected as the optimal platform operating parameter combination. This optimization process significantly improves the platform's operating efficiency and discrimination performance, ensuring stable and efficient operation of the platform under the constraints of edge hardware resources.

[0043] Preferably, the comprehensive decision expression for online determination of the maturity of camellia fruit is: ,in, For maturity category results, For class determination sign function, For the first The decision weights of the features in the pyramid layer. For the first The discriminative contribution value of the layer pyramid features. The number of characteristic layers of the pyramid. For process parameter correction coefficients, These are parameters used to determine the stability of the process.

[0044] Specifically, the comprehensive decision-making process for online maturity assessment of camellia oleifera fruit uses a category determination sign function to convert the weighted fusion feature values ​​into specific maturity category results. When the function input value is less than 0.3, it outputs an immature level; between 0.3 and 0.7, it outputs a mature level; and greater than 0.7, it outputs an over-ripe level. The decision weights of the t-th layer pyramid features are allocated based on the discriminative contribution of each layer's features, comprising four layers: layer 1 (bottom layer) with a decision weight of 0.15, layer 2 with 0.25, layer 3 with 0.3, and layer 4 (top layer) with 0.3, ensuring that key features at higher levels receive greater attention. The discriminative contribution value of the t-th layer pyramid features is obtained by calculating the average strength of maturity-related features in the feature graph of that layer, with a value ranging from 0 to 1. The number of pyramid feature layers is fixed at four, corresponding to the four-level structure of the feature pyramid network. The process parameter correction coefficient is set to 0.05 to correct systematic errors in the discrimination process. The discrimination process stability parameter is obtained by calculating the standard deviation of the discrimination contribution values ​​of the four-level features. A value of 1 is used when the standard deviation is less than 0.1, and a value of 0.5 is used when the standard deviation is greater than or equal to 0.1, quantifying the stability of the discrimination process. In the comprehensive decision-making process, the product of the decision weight and the discrimination contribution value of each level feature is first calculated. This product is then summed and added to the product of the process parameter correction coefficient and the discrimination process stability parameter to obtain the comprehensive decision value. Finally, the comprehensive decision value is mapped to a specific maturity category result using a category determination sign function. This decision rule fully integrates the discrimination information of multi-level features with process stability assessment, effectively improving the accuracy and reliability of maturity discrimination and reducing the impact of a single feature or an unstable discrimination process on the results.

[0045] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, starting the input layer of the deep residual camellia oleifera spectral analysis model, dividing the spectral feature data subset into multiple continuous feature segments according to wavelength intervals, with each feature segment corresponding to the initial input channel of the model; S32, performing preliminary extraction of features from each channel through the convolutional layer in the model to generate a basic spectral feature map, which is then transmitted to the first residual block for feature enhancement, retaining the original feature information through residual connections; S33, sequentially performing hierarchical iterative processing on the enhanced features through subsequent residual blocks, adjusting the feature distribution through batch normalization operations in each residual block, and enhancing the nonlinear expressive power of the features through activation functions; S34, performing global average pooling operations on the features output by the last residual block to obtain deep maturity association features with uniform dimensions, which are then transmitted to the feature fusion unit.

[0046] Specifically, the deep spectral feature extraction process in step S3 is refined, achieving accurate and efficient feature mining through four sub-steps. In S31, after activating the input layer of the deep residual camellia oleifera spectral analysis model, the spectral feature data subset divided in step S2 is further divided into six continuous feature segments according to wavelength ranges: 400-500 nm, 501-600 nm, 601-700 nm, 701-800 nm, 801-900 nm, and 901-1000 nm. Each feature segment corresponds to an initial input channel of the model, ensuring that the spectral data is input in an orderly manner according to wavelength bands. In S32, the 3×3 convolutional layer configured in the model performs preliminary feature extraction on each channel, with a convolution stride of 1 and a zero-padding parameter of 1, generating a 64×64-sized basic spectral feature map. This map is then transmitted to the first residual block, and the input features are directly superimposed onto the output of the residual block through residual connections, effectively preserving the original feature information and avoiding feature loss. S33 sequentially iterates through 11 residual blocks to process the enhanced features. Each residual block includes two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function layer. The momentum parameter of the batch normalization layer is set to 0.9, and the epsilon parameter is set to 1e-5. This operation adjusts the feature distribution, while the activation function enhances the non-linear expressive power of the features, enabling the model to capture more complex spectral correlation features. S34 converts the 128×128 feature map output from the last residual block into a 256-dimensional one-dimensional vector through global average pooling, i.e., a deep maturity correlation feature with uniform dimension. This feature is then transmitted to the feature fusion unit, laying the foundation for subsequent fusion with image features. The entire process, through hierarchical processing and residual mechanisms, fully mines the deep information related to the maturity of camellia fruit in the spectral data, improving feature quality.

[0047] Preferred, such as Figure 3 As shown, step S4 includes the following sub-steps: S41, decompose the feature data subset into feature dimensions to obtain multiple independent feature components such as color features, texture features, and morphological features, each feature component corresponding to a different correlation dimension of the maturity of camellia fruit; S42, calculate the correlation value between each feature component and the maturity category through the feature importance evaluation module in the adaptive weighted camellia maturity classification algorithm, and sort and select the core feature components based on the correlation value; S43, dynamically allocate the weight coefficient of each core feature component according to the selection result, assigning a higher weight coefficient to the feature component with a higher correlation value, while setting a weight adjustment threshold to limit the occurrence of extreme weights; S44, reorganize the weighted core feature components according to the preset dimensional order to construct a multi-dimensional image feature vector that can comprehensively reflect the maturity information of camellia fruit.

[0048] Specifically, the image feature processing flow in step S4 is broken down into four sub-steps: core feature selection, dynamic weight allocation, and feature vector construction. S41 first decomposes the image feature data subset divided in step S2 into feature dimensions. From the color feature data subset, eight independent feature components are extracted, including RGB three-channel pixel value distribution, hue mean, and saturation variance. From the texture feature data subset, twelve independent feature components are extracted, including gray-level co-occurrence matrix, texture entropy, and contrast. From the morphological feature data subset, eight independent feature components are extracted, including area, perimeter, roundness, and contour complexity. Each feature component corresponds to a different correlation dimension of camellia fruit maturity, comprehensively covering the key information in the image. S42 uses the feature importance evaluation module in the adaptive weighted camellia maturity classification algorithm. The Pearson correlation coefficient method is used to calculate the correlation value between each feature component and the maturity category, with a confidence level set at 95%. Based on the correlation values, 15 core feature components with correlation values ​​greater than 0.6 are selected, and redundant features with low correlation are removed, improving feature processing efficiency. S43 dynamically allocates the weight coefficients of each core feature component based on the screening results. Normalization is used to ensure the weight values ​​are within a reasonable range of 0.05 to 0.25. Feature components with higher correlation values ​​are assigned higher weight coefficients, with the highest correlation feature component having a weight of 0.25. A weight adjustment threshold of 0.05 is set to prevent the minimum weight from being lower than 0.05, ensuring the rationality of feature weight allocation and preventing individual feature components from being excessively ignored. S44 reorganizes the 15 weighted core feature components according to preset dimensions of color, texture, and morphology. During the reorganization process, Min-Max standardization is used to unify the numerical range of each feature component to the [0,1] interval, eliminating dimensional differences between different feature dimensions. Finally, a 15-dimensional multi-dimensional image feature vector that comprehensively reflects the maturity information of camellia fruit is constructed, providing high-quality image feature support for subsequent feature fusion.

[0049] Preferred, such as Figure 4As shown, step S5 includes the following sub-steps: S51, inputting deep spectral features and multi-dimensional image feature vectors into different input branches of the feature pyramid camellia oleifera maturity recognition network, and performing feature dimension matching processing through branch convolutional layers to ensure that the two types of features have the same dimensional specifications; S52, activating the network's multi-scale feature fusion module, performing downsampling and upsampling operations on the matched features at different scales to generate a set of multi-scale feature maps, with each scale feature map corresponding to maturity feature information at different levels; S53, stacking the multi-scale feature maps in hierarchical order through the pyramid hierarchical mapping unit, and performing information interaction and complementarity of features at different levels through a cross-layer connection mechanism to strengthen the expression of maturity features; S54, performing final convolutional compression processing on the stacked feature maps to generate a discriminative feature map that can be directly used for maturity discrimination, and transmitting it to the decision output unit at the edge.

[0050] Specifically, the multi-source feature fusion process in step S5 is refined, achieving efficient fusion of spectral features and image features and generation of discriminative feature maps through four sub-steps. In S51, the 256-dimensional deep spectral features obtained in step S3 and the 15-dimensional multi-dimensional image feature vector constructed in step S4 are respectively input into two independent input branches of the feature pyramid camellia oleifera maturity recognition network. The spectral feature branch compresses the feature dimension from 256 dimensions to 64 dimensions through a 1×1 convolutional layer, while the image feature branch expands the feature dimension from 15 dimensions to 64 dimensions through a 1×1 convolutional layer. The stride of the convolutional operation is set to 1 for both branches. The processing of the convolutional layer in this branch achieves dimensional matching of the two types of features, ensuring the feasibility of subsequent fusion. The S52 multi-scale feature fusion module of the network downsamples the matched 64-dimensional features using convolutional kernels of three different scales: 3×3, 5×5, and 7×7, generating feature maps of three sizes: 64×64, 32×32, and 16×16. Then, bilinear interpolation is used to upsample the 32×32 and 16×16 feature maps, restoring them to 64×64 size. These are then element-wise added to the original 64×64 feature map to generate a multi-scale feature map set, with each scale corresponding to different levels of maturity feature information. The S53 module divides the multi-scale feature maps into four levels according to priority using a pyramid hierarchical mapping unit, corresponding to basic features to core features from the bottom to the top. A cross-layer connection mechanism enables information interaction and complementarity between feature maps of different levels. Each feature map is then enhanced with two 3×3 convolutional layers, and an attention mechanism is introduced to highlight the response intensity of key feature regions. S54 performs final convolutional compression on the four stacked feature maps, using a 1×1 convolutional layer to compress the feature map dimension to a single channel, generating a 64×64 maturity discrimination feature map. Each pixel in this feature map corresponds to the maturity feature intensity of the target area of ​​the camellia fruit. The discrimination feature map is then transmitted to the decision output unit at the edge, providing accurate feature basis for subsequent maturity determination.

[0051] The formula in this invention integrates different scalar and vector parameters for unified calculation. Its core lies in constructing a standardized parameter interaction mechanism and dimension adaptation logic. Scalar parameters (such as environmental parameter correction coefficients and weight adaptive adjustment factors) are quantized and converted into numerical constants that can participate in calculations. Vector parameters (such as raw spectral wavelength data and image feature vectors) are processed through feature dimension regularization to form a fixed-dimensional vector structure. The formula design pre-defined parameter interface specifications, clearly defining the operational positions and interaction methods of scalars and vectors. For example, scalars are used as adjustment coefficients and multiplied element-wise with vectors, or summation operations are used to correct the overall characteristics of vectors using scalars. This ensures consistency in the operational dimensions of different types of parameters, laying the foundation for cross-type parameter fusion.

[0052] The key to the formula's collaborative calculation lies in unifying the physical meaning and mathematical properties of various parameters through mapping functions and operational rules. For the camellia fruit maturity assessment scenario, the formula incorporates specialized mapping functions (such as the spectral-environment mapping function) to transform parameters with different physical meanings (e.g., spectral data and environmental parameters) into quantitative indicators within the same feature space. Simultaneously, activation functions and similarity calculation functions are used to perform nonlinear transformations or normalization on the calculation results, eliminating dimensional and numerical range differences between parameters and enabling effective synergy between the scalar adjustment and vector feature representation. For example, the adaptive weighting algorithm uses exponential functions and summation operations to convert feature similarity (scalar) into weight coefficients, which are then weighted and fused with image feature vectors (vectors) to achieve an organic combination of different types of parameters.

[0053] The formula, through a hierarchical computational structure and constraint mechanism, ensures the rationality and effectiveness of calculations for different scalar and vector parameters. The computation process employs a hierarchical logic of "decomposition-fusion-correction," first extracting features and transforming dimensions of vector parameters, then performing targeted operations with scalar parameters, and finally optimizing the results through correction coefficients or constraint conditions. For example, in the feature pyramid network formula, spectral features and image features (vectors) are first processed through convolution and upsampling, then feature enhancement is achieved through element-wise multiplication of the feature enhancement matrix (vector), while parameter balancing coefficients (scalars) are used to adjust the weight ratios of different objective functions. Simultaneously, the formula incorporates regularization, weight thresholds, and other constraint mechanisms to prevent numerical anomalies or feature distortions when different types of parameters interact, ensuring the stability of cross-type parameter calculations and the adaptability of discrimination accuracy.

[0054] like Figure 5As shown, an online method for determining the maturity of camellia fruit based on deep learning edge computing is implemented through different units, including: a multi-source data collaborative acquisition unit, used to simultaneously acquire the original spectral data, spatial image data, and growth environment-related parameters of camellia fruit, and establish bidirectional data transmission connections with the spectral acquisition module, image acquisition module, and environmental sensors through a data interface; a feature data partitioning and processing unit, which receives the data output from the multi-source data collaborative acquisition unit, divides the spectral feature data subset and the image feature data subset according to data type, and transmits them to the corresponding feature processing units respectively; and a deep feature extraction and weighting unit, which incorporates a deep residual camellia spectral analysis model and an adaptive weighted camellia maturity classification algorithm. The system includes a method that receives feature data after partitioning, performs deep spectral feature extraction and image feature weighting, and outputs two types of feature vectors; a multi-scale feature fusion and recognition unit, equipped with a feature pyramid camellia oil maturity recognition network, receives feature vectors output by the deep feature extraction and weighting unit, and generates a discriminative feature map through multi-scale fusion and hierarchical mapping; a maturity decision output unit, based on the discriminative feature map, performs maturity category determination, outputs the determination result through a display interface, and establishes a connection with the data storage unit; and a full-process parameter storage unit, which receives raw data, intermediate feature data, and operating parameters transmitted by each functional unit, stores them according to timestamps and data types, and supports subsequent data retrieval and traceability.

[0055] This online method for determining the maturity of camellia oleifera fruit based on deep learning edge computing leverages the hierarchical extraction and residual connection mechanism of a deep residual model to deeply mine the deep features strongly correlated with maturity in spectral data, avoiding the bias caused by shallow features. An adaptive weighting algorithm dynamically evaluates the importance of each feature component in the image, assigning weights according to correlation and limiting extreme values ​​to ensure core features are fully highlighted. Furthermore, a feature pyramid network enables multi-scale feature fusion and cross-layer information interaction, comprehensively integrating key information from the spectrum and image. This effectively adapts to the feature differences of different growing environments and varieties of camellia oleifera fruit, significantly improving the comprehensiveness and accuracy of the determination criteria.

[0056] This method utilizes an integrated edge platform architecture to achieve local completion of the entire process of data acquisition, processing, and decision-making, avoiding latency and network dependence caused by cloud transmission and meeting the online discrimination requirements in field scenarios. The platform balances operational efficiency and discrimination performance through parameter optimization, avoiding edge-end lag caused by complex models without sacrificing discrimination accuracy. Each functional unit is efficiently connected according to the data flow, forming a closed-loop operation mechanism from multi-source data collaborative acquisition and partitioned processing to feature extraction, fusion recognition, and result output and parameter storage, solving the problem of poor connection between various links in traditional systems. At the same time, the full-process parameter storage function provides support for subsequent model iteration optimization and data traceability, further improving the practicality and scalability of the method.

[0057] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online determination of the maturity of camellia fruit based on deep learning edge computing, characterized in that, Includes the following steps: S1, through the multispectral acquisition module of the edge-end intelligent identification and analysis platform for camellia oleifera, acquires the original spectral data and spatial image data of the target area of ​​camellia oleifera fruit, and simultaneously collects the parameters related to the growth environment of camellia oleifera fruit; S2 transmits the acquired raw spectral data and spatial image data to the built-in data allocation unit at the edge, and divides the spectral feature data subset and the image feature data subset according to preset rules; S3, calls the deep residual camellia oleifera spectral analysis model to perform hierarchical feature extraction on the subset of spectral feature data, and mines the deep maturity correlation features in the spectral data through the residual block jump connection mechanism; S4. An adaptive weighted Camellia oleifera maturity classification algorithm is used to filter features and assign weights to a subset of image feature data to construct a multi-dimensional image feature vector. S5 inputs deep spectral features and weighted image feature vectors into the feature pyramid Camellia oleifera maturity recognition network, and generates maturity discrimination feature maps through multi-scale feature fusion and pyramid hierarchical mapping. The S6 edge-end intelligent identification and analysis platform for camellia oleifera outputs the maturity category results of camellia oleifera fruit based on the discrimination feature map, and stores the feature data and discrimination process parameters.

2. The method for online determination of Camellia oleifera fruit maturity based on deep learning edge computing according to claim 1, characterized in that, The expression for the deep residual spectral analysis model of Camellia oleifera is: ， in, To output eigenvalues ​​for spectral analysis, The original spectral wavelength parameters of camellia fruit, Optimize the parameter set for model training. For activation function, For the first Feature weight coefficients of each residual block For the first Each residual calculation module For environmental parameter correction factors, It is a spectrum-environment correlation mapping function. This is a set of parameters related to the growth environment of camellia oleifera fruit. This represents the number of residual blocks.

3. The method for online determination of Camellia oleifera fruit maturity based on deep learning edge computing according to claim 1, characterized in that, The expression for the adaptive weighted camellia oleifera maturity classification algorithm is as follows: ， in, The weighted image feature vector. The original image feature vector, This is the weight adaptive adjustment factor. This is the function for calculating feature similarity. This is the standard feature vector of maturity. The number of image feature dimensions. These are the feature gradient correction coefficients. This represents the gradient value of the eigenvector.

4. The method for online determination of camellia fruit maturity based on deep learning edge computing according to claim 1, characterized in that, The expression for the feature pyramid camellia oleifera maturity recognition network is as follows: ， in, For the first pyramid Layer fusion feature map, For deep spectral characteristics, For weighted image feature vectors, For convolution operations, These are the convolution kernel parameters for spectral features and image features, respectively. For upsampling operation, For feature element-wise multiplication operation For the first Layer feature enhancement matrix.

5. The method for online determination of camellia fruit maturity based on deep learning edge computing according to claim 1, characterized in that, The parameter optimization expression of the edge-end intelligent identification and analysis platform for camellia oil is: ， in, For the optimal combination of platform operating parameters, For platform parameter set, For maturity standard datasets, For parameter balance coefficients, For parameters The corresponding discrimination delay value, For parameters The deviation value between the discrimination result and the standard dataset.

6. The method for online determination of camellia fruit maturity based on deep learning edge computing according to claim 1, characterized in that, The comprehensive decision expression for online determination of the maturity of camellia fruit is as follows: ， in, For maturity category results, For class determination sign function, For the first The decision weights of the features in the pyramid layer. For the first The discriminative contribution value of the layer pyramid features. The number of characteristic layers of the pyramid. For process parameter correction coefficients, These are parameters used to determine the stability of the process.

7. The method for online determination of camellia fruit maturity based on deep learning edge computing according to claim 1, characterized in that, S3 includes the following steps: S31, start the input layer of the deep residual camellia oleifera spectral analysis model, divide the spectral feature data subset into multiple continuous feature segments according to wavelength range, and each feature segment corresponds to the initial input channel of the model; S32 extracts the features of each channel through the convolutional layer in the model to generate a basic spectral feature map, which is then transmitted to the first residual block for feature enhancement. The original feature information is preserved through residual connections. S33, the enhanced features are processed hierarchically through subsequent residual blocks. Each residual block adjusts the feature distribution through batch normalization and enhances the nonlinear expressive power of the features through activation functions. S34. The features output by the last residual block are subjected to global average pooling to obtain deep maturity association features with uniform dimensions, which are then transmitted to the feature fusion unit.

8. The method for online determination of Camellia oleifera fruit maturity based on deep learning edge computing according to claim 1, characterized in that, S4 includes the following steps: S41, the feature dimension of the image feature data subset is decomposed to obtain multiple independent feature components such as color feature, texture feature, and morphology feature. Each feature component corresponds to a different correlation dimension of the ripeness of the camellia fruit. S42, through the feature importance evaluation module in the adaptive weighted camellia oleifera maturity classification algorithm, calculate the correlation value between each feature component and the maturity category, and sort and select the core feature components based on the correlation value; S43: Dynamically allocate the weight coefficients of each core feature component based on the screening results. Feature components with higher correlation values ​​are assigned higher weight coefficients. At the same time, a weight adjustment threshold is set to limit the occurrence of extreme weights. S44. The weighted core feature components are reorganized according to the preset dimension order to construct a multi-dimensional image feature vector that can comprehensively reflect the maturity information of camellia fruit.

9. The method for online determination of camellia fruit maturity based on deep learning edge computing according to claim 1, characterized in that, S5 includes the following steps: S51, deep spectral features and multi-dimensional image feature vectors are respectively input into different input branches of the feature pyramid camellia oleifera maturity recognition network, and feature dimension matching is performed through branch convolutional layers to make the two types of features have the same dimensional specifications. S52, start the network's multi-scale feature fusion module, perform downsampling and upsampling operations on the matched features at different scales, and generate a set of multi-scale feature maps, with each scale feature map corresponding to maturity feature information at different levels; S53 uses a pyramid-level mapping unit to stack multi-scale feature maps in a hierarchical order, and uses a cross-layer connection mechanism to enable information exchange and complementarity of features at different levels, thereby strengthening the expression of maturity features. S54 performs final convolutional compression on the superimposed feature maps to generate a discriminative feature map that can be directly used for maturity determination, and transmits it to the decision output unit at the edge.

10. The method for online determination of Camellia oleifera fruit maturity based on deep learning edge computing according to any one of claims 1-9, characterized in that, This method is implemented through different units, including: The multi-source data collaborative acquisition unit is used to simultaneously acquire the original spectral data, spatial image data and growth environment-related parameters of camellia fruit, and establishes a two-way data transmission connection with the spectral acquisition module, image acquisition module and environmental sensor through the data interface; The feature data partitioning processing unit receives the data output by the multi-source data collaborative acquisition unit, divides the spectral feature data subset and the image feature data subset according to the data type, and transmits them to the corresponding feature processing units respectively. The deep feature extraction and weighting unit has a built-in deep residual Camellia oleifera spectral analysis model and an adaptive weighted Camellia oleifera maturity classification algorithm. It receives feature data after partitioning and completes deep spectral feature extraction and image feature weighting processing, and outputs two types of feature vectors. The multi-scale feature fusion recognition unit is equipped with a feature pyramid camellia oil maturity recognition network. It receives the feature vectors output by the deep feature extraction and weighting unit, and generates a discriminative feature map through multi-scale fusion and hierarchical mapping. The maturity decision output unit performs maturity category determination based on the discriminative feature map, outputs the determination result through the display interface, and establishes a connection with the data storage unit. The end-to-end parameter storage unit receives raw data, intermediate feature data, and operating parameters transmitted from various functional units, and stores them according to timestamps and data types, supporting subsequent data retrieval and traceability.

Images