Method and system for detecting white blight of morel

Abstract

The present application relates to a kind of Morel planting intelligent control method and system, by dividing Morel growth into mycelium period, primordium period and fruit body period three stages, based on multi-source sensing and visual data Construction environment and growth dual-dimensional feature dataset, establish different deep reinforcement learning model.For different growth stages Set hierarchical reward function, combined with optimized Actor-Critic architecture, stage priority experience replay and policy smoothing regularization mechanism, realize the adaptive optimization of control strategy.By introducing environmental disturbance adaptive fine-tuning and closed-loop feedback control, realize the precise collaborative control of temperature and humidity, illumination and gas concentration in shelter.The present application can significantly improve Morel yield and quality, reduce the intensity of manual intervention, improve system stability and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of morel disease control and edge intelligent detection technology, specifically to a lightweight time-series image anomaly detection device for early latent characteristics of white mold in morel container planting scenarios, as well as a supporting low-power edge node deployment architecture and early warning method. Background Technology

[0002] Morel white mold is a core disease restricting large-scale cultivation in mobile cabins. In its early stages, the disease only manifests as subtle features such as small white spots on the surface of the culture medium and abnormal mycelial flocculent structures, which are difficult to identify with the naked eye. Traditional manual inspections suffer from high latency and a high rate of missed detection, which can easily lead to large-scale spread of the disease. Existing disease detection technologies mostly use cloud-based deep learning models, which suffer from high data transmission latency, high cloud computing costs, and high network dependence. Conventional edge detection models have redundant parameters and high computing power requirements, which cannot be adapted to the low-power, low-computing-power hardware conditions of embedded edge nodes in mobile cabins, making it difficult to achieve real-time early warning of white mold. Summary of the Invention

[0003] In view of the problems existing in the prior art, the present invention provides a method for detecting morel white mold disease, the method comprising the following steps: Step S1: Deploy cameras in the morel mushroom cultivation area on the modular shelter shelf and continuously collect time-series images at a fixed frequency; Step S2: Preprocess the time series image; Step S3: Burn the trained lightweight temporal anomaly detection model to the edge computing node. The model quickly determines whether there are early anomalies of white mold based on the extracted temporal features. The quantization inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms. Step S4: When the detection is normal, the edge node only uploads status data; when it is determined to be suspected early white mold, a local audible and visual warning is immediately triggered, and the abnormal time series image and detection results are simultaneously uploaded to the cloud platform to facilitate quick handling by growers; Step S5: The cloud platform collects the abnormal samples uploaded by the edge nodes, optimizes the model through incremental learning, and then sends the lightweight update package to the edge nodes. Preferably, in step 2, the temporal image is preprocessed by grayscale conversion, Gaussian filtering, and ROI cropping to reduce image noise and data volume; by improving the GhostNet lightweight network, latent features such as texture changes and abnormal colony morphology between temporal frames are extracted to avoid misjudgment problems in single-frame image detection. Preferably, in step 4, the edge node supports offline detection and can still provide normal early warnings even without a network, ensuring adaptability to planting scenarios. Preferably, the improved GhostNet lightweight network includes: Temporal Channel Stitching Layer: Used to stitch together T consecutively acquired grayscale temporal images along the channel dimension to construct... The initial temporal feature tensor; where T is an integer from 3 to 5; Temporal Difference Ghost Module: Includes standard convolutional layers and temporal difference operation layers; the standard convolutional layers generate a sequence of basic feature maps, and the temporal difference operation layers calculate the difference terms of feature maps at adjacent time steps and perform weighted fusion with the basic feature maps to generate a phantom feature map with extracted colony dynamic growth features; Microtexture Attention Module: Embedded after the temporal difference Ghost module, it uses the high-pass filtering principle to extract high-frequency edge information in the feature map, and generates a spatial attention mask through global max pooling and local variance pooling to enhance the expression of the flocculent latent features at the base of morel mushrooms. The present invention also provides a morel white mold detection system, the system comprising: a time-series image acquisition unit, an image processing unit, an edge computing node unit, a data transmission unit, a local early warning unit, and a cloud platform; The time-series image acquisition unit is used to acquire continuous frame images of morel mushrooms inside the shelter. The image processing unit preprocesses the time-series image; The edge computing node unit burns the trained lightweight temporal anomaly detection model into the edge computing node unit. The model quickly determines whether there are early anomalies of white mold based on the extracted temporal features. The quantization inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms. When the edge computing node unit determines that the detection is in a normal state, the edge node unit only uploads the status data; when the edge computing node unit determines that it is suspected to be early white mold, it immediately triggers a local audible and visual warning, and the data transmission unit simultaneously uploads the abnormal time sequence image and detection results to the cloud platform. When the edge computing node unit determines that the disease is suspected to be early-stage white mold, the local early warning unit immediately triggers a local audible and visual early warning. The cloud platform collects abnormal samples uploaded by the data transmission unit, optimizes the model through incremental learning, and then sends the lightweight update package to the edge nodes. Preferably, the image processing unit performs grayscale conversion, Gaussian filtering, and ROI cropping preprocessing on the temporal image in sequence to reduce image noise and data volume; by improving the GhostNet lightweight network, it extracts latent features such as texture changes and abnormal colony morphology between temporal frames to avoid misjudgment problems in single-frame image detection. Preferably, the edge node supports offline detection and can still provide normal early warnings even without a network, ensuring adaptability to planting scenarios. Preferably, the edge computing node unit adopts a master-slave heterogeneous architecture and realizes data transmission through a high-speed SPI interface in conjunction with a DMA controller; and the edge computing node unit is equipped with two sets of independent pixel buffers, which utilize a ping-pong operation mechanism to realize the parallel reception of the next frame of time-series image data while the embedded AI acceleration chip is performing the current frame inference. Preferably, the system further includes a power supply unit; the power supply unit provides power to the time-series image acquisition unit, the edge computing node unit, the local early warning unit, and the data transmission unit. Preferably, the time-series image acquisition unit further includes an environment adaptive calibration module to avoid a decrease in the recognition accuracy of the general network in the mobile cabin planting scenario, and to ensure that the detection network is highly adapted to the detection needs and planting scenarios of morel white mold, and is specifically used for early warning of morel white mold. The innovative aspects of the embodiments in this specification include: 1. This invention constructs a three-in-one white mold detection architecture that integrates "temporal texture feature extraction - lightweight anomaly discrimination - edge inference acceleration", designs a lightweight temporal anomaly detection network based on GhostNet improvement, and combines it with an edge node hardware-algorithm co-optimization mechanism to break through the bottleneck of computing power and power consumption at the edge. 2. Unlike existing technologies that use large-volume convolutional networks and single-frame image detection, this invention targets the temporal spread characteristics of morel white mold, extracts texture change features from consecutive frame images, and compresses the model size to less than 5MB through model pruning, depthwise separable convolution, and quantized perceptual training. At the same time, it optimizes the hardware scheduling logic of edge nodes to achieve millisecond-level inference on low-power embedded devices, balancing lightweight design and detection accuracy, and filling the technical gap in early warning of morel white mold at the edge. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments or related technologies of this specification, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the structure of the morel white mold detection system provided by the present invention; Figure 2a Images were collected for areas prone to white mold, such as morel mycelium and the base of fruiting bodies. Figure 2b Post-processed images were collected for areas prone to white mold, such as morel mycelium and the base of fruiting bodies. Figure 3 The flowchart is a method for detecting morel white mold disease provided by the present invention. Reference numerals: 10, image acquisition unit; 20, image processing unit; 30, edge computing node unit; 40, local early warning unit; 50, data transmission unit; 60, power supply unit; 100, detection system. Detailed Implementation The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this specification are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Example 1 like Figure 1 The diagram shown is a structural schematic of the morel white mold detection system provided by the present invention. The morel white mold detection system 100 includes a time-series image acquisition unit 10, an image processing unit 20, an edge computing node unit 30, a local early warning unit 40, a data transmission unit 50, and a power supply unit 60. The time-series image acquisition unit 10 is used to acquire continuous frame images of morel mushrooms inside the container; specifically, it involves deploying miniature high-definition cameras in the morel mushroom cultivation area on the shelf of the container, continuously acquiring 3-5 frames of time-series images at a fixed frequency, focusing on areas prone to white mold disease such as the surface of the culture medium, morel mycelium, and the base of the fruiting bodies (e.g., the surface of the culture medium, the base of the fruiting bodies, etc.). Figure 2a As shown), remove blurry, overexposed, and invalid frames (such as...). Figure 2b As shown in the figure, a continuous and valid time-series image sequence is preserved. Experimental data: The camera capture resolution was 1080P, the capture frequency was 1 time / 10min, the effective frame retention rate reached 95.6%, and the invalid frames were mainly blurry frames caused by airflow in the cabin, accounting for only 4.4%. The image processing unit 20 performs grayscale conversion, Gaussian filtering, and ROI cropping preprocessing on the acquired temporal images to reduce image noise and data volume. By improving the GhostNet lightweight network, it extracts latent features such as texture changes and abnormal colony morphology between temporal frames, avoiding misjudgments in single-frame image detection. Experimental data: After preprocessing, image noise is reduced by 68.9%, and data volume is compressed by 72%; temporal feature extraction time is ≤30ms / frame, and the misjudgment rate is reduced by 53.8% compared to single-frame image feature extraction. In a preferred embodiment, to achieve lightweight deployment on the edge computing nodes of the mobile cabin and accurately capture the subtle temporal morphological changes in the early stages of morel white mold disease, this embodiment improves the standard GhostNet network specifically for temporal feature extraction, constructing a lightweight temporal anomaly detection network that includes a "temporal channel splicing input layer," a "temporal difference Ghost module (TD-GhostModule)," and a "microtexture attention mechanism." The specific structure and data processing flow are as follows: Unlike the traditional three-channel (RGB) input of a single-frame image, the network of this invention first integrates continuously acquired valid frames through a temporal channel stitching layer. Let the continuous T-frame image sequence output by the image processing unit after preprocessing be... (This embodiment is preferred) ), where the size of a single frame image is (C=1, as grayscale has been performed during preprocessing). The input layer physically stitches these T frames of images along the channel dimension to form a shape of size [size missing]. The initial temporal feature tensor. This concatenation method avoids the introduction of computationally complex 3D convolution (3D-CNN), and provides the original temporal span information for subsequent networks with almost no increase in the number of network parameters. The original GhostNet generates intrinsic feature maps through a small number of standard convolutions, and then generates phantom feature maps using simple linear operations (CheapOperations). To enhance the capture of the dynamic growth process of "flocculent abnormality" in the early stage of white mold, this invention replaces the original linear operations with cross-channel "temporal difference operations" to construct the TD-Ghost module. Specifically, in the TD-Ghost module, assuming the first step is conventional... The sequence of feature maps generated by convolution is aligned by time steps. During the phantom feature map generation stage, the network not only analyzes the features of the current frame... Perform depthwise separable convolution operations Simultaneously, differential calculation of adjacent frames is introduced to extract dynamic texture change features. The formula for calculating the difference feature is as follows: ⊕ Where ⊕ represents the feature fusion operation, These are the differential weight coefficients that can be learned through backpropagation. By introducing... The TD-Ghost module can force the network to learn the subtle expansion of colony area and the gradual change in white spot concentration, thereby effectively avoiding the problem of misjudgment of white spots caused by light reflection in a single frame image. In this embodiment, the TD-Ghost module employs a depthwise separable convolution operation. It is the core method for achieving lightweight feature extraction. Unlike traditional standard convolution, the operation... The convolution process is broken down into two independent steps: "depthmarch" and "pointwise convolution," thereby significantly reducing computational cost while maintaining feature extraction capabilities. 1. Depthwise convolution stage: When processing feature tensors that have undergone temporal concatenation, the operation First, a convolutional kernel is applied independently to each channel of the input. Let the number of channels in the input feature map be M, then M kernels of size M are set accordingly. The filters are configured such that each filter performs spatial convolution on only one corresponding channel, without cross-channel feature fusion. This step is only responsible for extracting spatial texture features within each temporal frame, and its computational cost is only 1 / M of that of traditional convolution. 2. Pointwise convolution stage: After depthwise convolution, in order to enable information exchange between channels (i.e., between different time steps), operations are performed... Next, pointwise convolution is performed. The computation logic is as follows: using a size of... The convolutional kernels linearly combine the feature maps output from the previous stage across channels. This step is responsible for mapping the spatial features extracted by the depthwise convolution to a high-dimensional feature space, completing the fusion and dimensionality reduction of temporal features, and ensuring that the model can learn the evolution of pixel values ​​caused by disease growth between different frames. Let the number of input channels be M, the number of output channels be N, and the kernel size be M. Standard convolution computational cost: Depthwise separable convolution Computational complexity: Gain Ratio: In this system (typically k=3, N is much greater than 1), the operation... The number of parameters and computational cost are approximately 1 / 8 to 1 / 9 of that of standard convolution. This structural redundancy removal forms the basis for the algorithms described in this invention, which aim to "compress the inference time of a single frame image to within 50ms" and "make the model size less than 5MB." This is in conjunction with the aforementioned temporal difference term. ,operate This allows the network to locate tiny white, flocculent anomalies at the base of morel mushrooms with extremely low computing power. Morel mushrooms have a complex network-like folded structure (normal morphology), which can easily be confused with the early characteristics of white mold disease. Therefore, in the deep feature extraction stage of the network, this embodiment embeds a lightweight microtexture attention mechanism after the TD-Ghost module. This mechanism includes a pooling layer based on high-pass filtering, which focuses on extracting high-frequency edge information from the feature map. Let the input feature map be... Attention mask The generation process is as follows: First, for Global max pooling and local variance pooling are performed separately (to highlight texture irregularity). The concatenation of these two methods is then passed through a shared multilayer perceptron (MLP) for dimensionality reduction and activated using a sigmoid function to obtain spatial attention weights. : ⊕ ; In the above formula, Input feature map: This represents the original feature tensor entering the attention module. In this embodiment, it is the temporal feature map extracted by the TD-Ghost module, containing the spatial and temporal features of the initially captured morel growth region. MaxPool() represents the max pooling operation: used for feature map extraction. The maximum pixel response value within a local window or global scope. The main function of this operation is to extract the most prominent texture features (such as bright spots of hyphae) in the background and diseased areas, while preserving high-frequency signals. VarPool() represents the variance pooling operation: the core feature extraction operator of this invention, used to calculate the pixel variance within a local feature region. Since the normal fruiting body folds of morel mushrooms typically have a relatively regular texture, while early white mold hyphae exhibit irregular, flocculent, splattered characteristics, variance pooling can effectively quantify the irregularity (i.e., the degree of dispersion) of this texture, thereby highlighting suspected lesion areas. MLP() stands for Multilayer Perceptron: It represents a fully connected neural network layer with shared parameters, used to perform nonlinear transformations and dimensionality compression on pooled features. Its function is to learn the correlation between different feature channels and automatically select the feature weights that contribute the most to the identification of white mold. ⊕ indicates element-level addition / feature fusion: This means that the feature vectors processed by the max pooling path and the variance pooling path are fused by element-level addition, so that the final attention mask can simultaneously take into account salient features (brightness) and texture complexity (variance). This refers to the Sigmoid activation function, which maps the fused features to the (0, 1) interval. Its output value is the weight coefficient; the closer the value is to 1, the higher the probability that the region belongs to the abnormal features of white mold. The attention mask represents the final generated spatial attention weight matrix. The system uses this matrix to... Compared with the original feature map Element-wise multiplication is performed to enhance abnormal texture regions while suppressing morel background noise. Final output features ⊕ This mechanism enables the network to automatically suppress the macroscopic structural features of morel fruiting bodies during feature selection, focusing attention on the abnormal flocculent high-frequency textures on the surface and base of the culture medium, greatly improving the specificity of disease identification. By introducing an attention mechanism based on variance pooling (VarPool), this invention cleverly utilizes the significant difference in spatial distribution entropy (irregularity) between normal morel tissue and white mold hyphae, solving the problem that traditional attention mechanisms (relying only on max pooling or average pooling) are prone to misjudging background wrinkles in low-contrast environments, and providing algorithmic support for achieving an early recognition rate of over 97%. After passing through multiple TD-Ghost modules and global average pooling, the network outputs a binary classification result (normal / suspected white mold) through a fully connected layer. To meet the stringent requirement of reducing single-frame image inference time to less than 50ms for edge nodes, Quantization-Aware Training (QAT) was employed during the cloud training phase. INT8 quantization error was simulated during the forward propagation of training, and the weights were fine-tuned to adapt to low-precision computation. After training, the network parameters (weights and activation values) were statically quantized from 32-bit floating-point numbers (FP32) to 8-bit fixed-point numbers (INT8). Due to the low-parameter characteristics of the TD-Ghost module itself combined with INT8 quantization, the final model size burned to the edge nodes was successfully compressed to 4.8MB, achieving millisecond-level fast inference on the STM32 hardware platform. The edge computing node unit 30 adopts a hardware architecture of low-power STM32 + embedded AI acceleration chip, suitable for the closed planting environment of the container without strong power supply. The trained lightweight temporal anomaly detection model is burned into the edge computing node. Based on the extracted temporal features, the model quickly determines whether there are early anomalies of white mold. Quantized inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms, ensuring real-time detection. Experimental data: The model size is 4.8MB, and the single-frame inference time at the edge is 42ms, which is 59.4% higher than the unaccelerated lightweight model; the accuracy rate of early white mold (1-2 days after onset) identification reaches 96.7%, with a false negative rate of only 2.3% and a false positive rate of 1.0%. In a preferred embodiment, the task scheduler built into the edge computing node unit 30 employs a priority interrupt mechanism. The image acquisition unit 10 stores the acquired raw time-series frames into an external SRAM cache via DMA (Direct Memory Access), without CPU intervention, thus reducing power consumption. The task scheduler immediately wakes up the embedded AI acceleration chip after detecting that the cache is full of 5 frames. The software layer pushes the INT8 quantized model weights and temporal feature tensors to the acceleration chip's computing array via the HAL (Hardware Abstraction Layer) interface. During non-sampling periods (e.g., a 10-minute interval), the system is forced into a "deep sleep" mode, shutting down the camera and acceleration chip power, retaining only the RTC (Real-Time Clock) to trigger the next sampling. Experimental data demonstrates that this software-hardware co-scheduling reduces the system's average power consumption by more than 65%. In a preferred embodiment, the STM32 general-purpose processor (as the main controller) and the embedded AI acceleration chip are physically connected via a high-speed synchronous serial interface SPI (Serial Peripheral Interface), and dual-SPI or quad-SPI mode is enabled, with a communication rate configured to be no less than 40MHz. The STM32 microcontroller sends model start commands, weight quantization parameters, and ROI clipping coordinates to the accelerator chip via the standard SPI protocol. To overcome the bottleneck of serial communication, the system adopts DMA (Direct Memory Access) mode. After the image processing unit 20 completes preprocessing, the STM32, without CPU intervention, directly pushes pixel data in batches to the external bus interface of the accelerator chip through the DMA controller, greatly reducing CPU usage and latency caused by data transfer. To eliminate the waiting time between image acquisition and inference, the edge computing node allocates two independent cache spaces in SRAM, namely ping-pong cache region A and region B: While the accelerator chip is extracting and inferring the timing features of the current frame (the nth frame) in buffer area A, the STM32 master controller synchronously writes the preprocessed data of the next frame (the n+1th frame) into buffer area B via DMA. Once inference in region A is complete and data transfer in region B is finished, the system instantly switches pointers via a hardware semaphore, causing the accelerator chip to switch to processing region B, while region A is released to receive the (n+2)th frame of data. This "processing and data transfer synchronization" mechanism ensures that the AI ​​accelerator chip is always operating at full capacity, effectively keeping the total system latency within the single-frame inference time (approximately 35ms-42ms). The accelerator chip employs memory-mapped I / O technology, directly mapping its internal registers and computing arrays to the STM32's address space. Simultaneously, for the INT8 quantization model, the system compresses image pixels from 32-bit floating-point numbers to 8-bit fixed-point numbers, increasing the information density of a single DMA transfer by four times. Through the collaboration of the aforementioned high-speed interface and dual-buffer management mechanism, the system successfully avoids the computational bottleneck caused by bus contention and serial waiting. The actual measured latency of the entire machine from the end of preprocessing to the output of detection results for a single frame image is stable within 45ms, which meets the real-time monitoring requirements for early abnormalities of white mold in the mobile cabin environment. The cloud platform collects abnormal samples uploaded by the data transmission unit, optimizes the model through incremental learning, and then distributes lightweight update packages to edge nodes. This continuously improves detection accuracy without requiring hardware replacement, adapting to the detection needs of different strains and mobile testing environments. Experimental data: For every 50 additional abnormal samples, the model's recognition accuracy improves by 1.2%-1.8%; after three iterations, the early identification accuracy for different morel strains of white mold remains stable above 97%, with a 47.9% improvement in adaptability. When the edge computing node unit determines that the detection is in a normal state, the data transmission unit 50 only uploads the status data; when the edge computing node unit determines that it is suspected to be early white mold, it immediately triggers a local audible and visual warning, and the data transmission unit simultaneously uploads the abnormal time sequence image and detection results to the cloud management platform to facilitate quick handling by planting personnel; the cloud data upload delay is ≤5s. The local early warning unit 40 immediately triggers a local audible and visual early warning when the edge computing node unit determines that early white mold is suspected. The edge computing node unit supports offline detection and can still issue warnings normally when there is no network, ensuring adaptability to planting scenarios. Experimental data: The local early warning response time is ≤100ms. In offline mode, the edge node can continuously issue warnings for 72 hours (low power mode) without any decrease in warning accuracy. The power supply unit 60 provides power to the time-series image acquisition unit, edge computing node unit, local early warning unit, and data transmission unit. The lightweight temporal anomaly detection network used in this system is not a general-purpose image detection network. Instead, it has been specifically improved for the early characteristics of morel white mold disease and the cultivation scenario. It is a customized design specifically for morel mushrooms: First, for the latent characteristics of early morel white mold disease, which only manifest as small white spots on the culture medium and mycelial flocculent abnormalities, the feature extraction layer of the network is optimized to enhance the feature capture ability of subtle texture changes and low-contrast areas, avoiding the problem of general networks missing detection of this latent disease in morel mushrooms. Second, in combination with the morphological differences and spectral characteristics of mycelia, fruiting bodies and white mold colonies in morel mushroom container cultivation, feature screening rules are customized to eliminate interference from the normal growth form of morel mushrooms and improve the specificity of disease identification. Third, it is adapted to environmental changes during the growth cycle of morel mushrooms, such as image grayscale shifts caused by temperature and humidity fluctuations in the container. An environmental adaptive calibration module is added to avoid the decrease in recognition accuracy of general networks in the container cultivation scenario, ensuring that the detection network is highly adapted to the detection needs and cultivation scenario of morel white mold disease. It is specifically designed for early warning of morel white mold disease, rather than general disease detection. Experimental data: After the dedicated improvements, the network achieved a capture rate of 97.8% for early latent features of morel white mold disease (white spots diameter ≥0.2mm), which is 41.3% higher than the general detection network; the accuracy of removing interference from the normal growth morphology of morels reached 99.1%; when the environmental grayscale shift was ±15%, the recognition accuracy decreased by ≤3.2%, which is far better than the general network (decrease ≥18.7%). Example 2 like Figure 3 The diagram shows a flowchart of the method for detecting morel white mold provided by the present invention. The method includes the following steps: Step S1: Deploy miniature high-definition cameras in the morel mushroom cultivation area of ​​the modular container, continuously acquire 3-5 frames of time-series images at a fixed frequency, focus on the surface of the culture medium, morel mycelium and the base of the fruiting body, and other areas prone to white mold disease, remove blurry and overexposed invalid frames, and retain the continuous and valid time-series image sequence. Step S2: Perform grayscale conversion, Gaussian filtering, and ROI cropping preprocessing on the acquired time-series images to reduce image noise and data volume; by improving the GhostNet lightweight network, extract latent features such as texture changes and abnormal colony morphology between time-series frames to avoid misjudgment problems in single-frame image detection. Step S3: Burn the trained lightweight temporal anomaly detection model to the edge computing node. Based on the extracted temporal features, the model quickly determines whether there are early anomalies of white mold. The quantization inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms to ensure real-time detection. Step S4: When the detection is normal, the edge node only uploads status data; when it is determined to be suspected early white mold, a local audible and visual warning is immediately triggered, and the abnormal time sequence image and detection results are uploaded to the cloud management platform at the same time, so that the growers can deal with it quickly; the edge node supports offline detection and can still issue warnings normally when there is no network, ensuring the adaptability of the planting scenario. Step S5: The cloud platform collects abnormal samples uploaded by edge nodes, optimizes the model through incremental learning, and then distributes the lightweight update package to the edge nodes. This continuously improves detection accuracy without replacing hardware and adapts to the detection needs of different strains and different mobile cabin environments. Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention. Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting morel white mold disease, characterized in that, The method includes the following steps: Step S1: Deploy cameras in the morel mushroom cultivation area on the modular shelter shelf and continuously collect time-series images at a fixed frequency; Step S2: Preprocess the time series image; Step S3: Burn the trained lightweight temporal anomaly detection model to the edge computing node. The model quickly determines whether there are early anomalies of white mold based on the extracted temporal features. The quantization inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms. Step S4: When the detection is normal, the edge node only uploads status data; when it is determined to be suspected early white mold, a local audible and visual warning is immediately triggered, and the abnormal time series image and detection results are simultaneously uploaded to the cloud platform to facilitate quick handling by growers; Step S5: The cloud platform collects the abnormal samples uploaded by the edge nodes, optimizes the model through incremental learning, and then sends the lightweight update package to the edge nodes.

2. The detection method according to claim 1, characterized in that, In step 2, the temporal images are preprocessed sequentially by grayscale conversion, Gaussian filtering, and ROI cropping to reduce image noise and data volume. By improving the GhostNet lightweight network, latent features such as texture changes and abnormal colony morphology between temporal frames are extracted to avoid misjudgment problems in single-frame image detection.

3. The detection method according to claim 1, characterized in that, In step 4, the edge node supports offline detection and can still provide warnings even without a network, ensuring adaptability to planting scenarios.

4. The detection method according to claim 1, characterized in that, The improved GhostNet lightweight network includes: Temporal Channel Stitching Layer: Used to stitch together T consecutively acquired grayscale temporal images along the channel dimension to construct... The initial temporal feature tensor; where T is an integer from 3 to 5; Temporal Difference Ghost Module: Includes standard convolutional layers and temporal difference operation layers; the standard convolutional layers generate a sequence of basic feature maps, and the temporal difference operation layers calculate the difference terms of feature maps at adjacent time steps and perform weighted fusion with the basic feature maps to generate a phantom feature map with extracted colony dynamic growth features; Microtexture Attention Module: Embedded after the temporal difference Ghost module, it uses the high-pass filtering principle to extract high-frequency edge information in the feature map, and generates a spatial attention mask through global max pooling and local variance pooling to enhance the expression of the flocculent latent features at the base of morel mushrooms.

5. A morel white mold detection system, characterized in that, The system includes: a time-series image acquisition unit, an image processing unit, an edge computing node unit, a data transmission unit, a local early warning unit, and a cloud platform; The time-series image acquisition unit is used to acquire continuous frame images of morel mushrooms inside the shelter. The image processing unit preprocesses the time-series image; The edge computing node unit burns the trained lightweight temporal anomaly detection model into the edge computing node unit. The model quickly determines whether there are early anomalies of white mold based on the extracted temporal features. The quantization inference acceleration technology is used to compress the inference time of a single frame image to less than 50ms. When the edge computing node unit determines that the detection is in a normal state, the edge node unit only uploads the status data; when the edge computing node unit determines that it is suspected to be early white mold, it immediately triggers a local audible and visual warning, and the data transmission unit simultaneously uploads the abnormal time sequence image and detection results to the cloud platform. When the edge computing node unit determines that the disease is suspected to be early-stage white mold, the local early warning unit immediately triggers a local audible and visual early warning. The cloud platform collects abnormal samples uploaded by the data transmission unit, optimizes the model through incremental learning, and then sends the lightweight update package to the edge nodes.

6. The system according to claim 5, characterized in that, The image processing unit performs grayscale conversion, Gaussian filtering, and ROI cropping preprocessing on the time-series images in sequence to reduce image noise and data volume. By improving the GhostNet lightweight network, it extracts latent features such as texture changes and abnormal colony morphology between time-series frames to avoid misjudgment problems in single-frame image detection.

7. The detection system according to claim 5, characterized in that, The edge nodes support offline detection and can still provide warnings even without a network, ensuring adaptability to various planting scenarios.

8. The detection system according to claim 5, characterized in that, The edge computing node unit adopts a master-slave heterogeneous architecture and realizes data transmission through a high-speed SPI interface in conjunction with a DMA controller. The edge computing node unit is equipped with two sets of independent pixel buffers, and uses a ping-pong operation mechanism to realize the parallel reception of the next frame of time sequence image data while the embedded AI acceleration chip is performing the current frame inference.

9. The detection system according to claim 5, characterized in that, The system also includes a power supply unit; the power supply unit provides power to the time-series image acquisition unit, the edge computing node unit, the local early warning unit, and the data transmission unit.

10. The detection system according to claim 5, characterized in that, The time-series image acquisition unit also includes an environment adaptive calibration module to avoid a decrease in the recognition accuracy of the general network in the mobile cabin planting scenario, and to ensure that the detection network is highly adapted to the detection needs and planting scenarios of morel white mold, and is specifically used for early warning of morel white mold.

Images