Large model-based algae multi-dimensional feature intelligent identification system and method

Abstract

The invention belongs to the technical field of water environment algae monitoring, and provides an algae multi-dimensional feature intelligent identification system and method based on a large model. Comprising a data acquisition module, a preprocessing module, a feature extraction module, a feature library module, a large model reasoning module, a result verification module, an incremental training module and a real-time monitoring terminal module. According to the method, the problems of efficiency, precision, adaptability, updating capability and the like of a traditional algae recognition technology are effectively solved, the algae recognition efficiency and precision are remarkably improved, the scene adaptability and the dynamic updating capability are enhanced, and the method has wide application prospects in the fields of water environment monitoring and the like.

Description

Technical Field

[0001] This invention belongs to the field of aquatic environment algae monitoring technology, and particularly relates to an intelligent identification system and method for multi-dimensional algae features based on a large model. Background Technology

[0002] In the field of algae identification, current technologies have many limitations, which can be summarized into four aspects: efficiency, accuracy, adaptability, and update capability.

[0003] In terms of efficiency: Manual identification relies on professionals using microscopes to determine the species of algae based on characteristics such as algal cell structure, pigment distribution, and colony morphology, in accordance with professional literature. This method is extremely inefficient, with a single sample processing time ranging from 5 to 30 minutes. In large-scale water body monitoring, facing tens of thousands of samples, it is difficult to meet the timeliness requirements. For example, in watershed water quality surveys, manual identification requires a large amount of manpower and time.

[0004] In terms of accuracy: Human identification is highly subjective. For similar algae with subtle morphological differences, such as diatoms from different genera whose shell stripe density differs by only 1-2 stripes / μm, expert judgment is easily influenced by personal experience, fatigue, and emotions, resulting in an accuracy rate fluctuating between 60% and 85%. Basic image recognition technologies, which employ models such as support vector machines and shallow neural networks, rely on manual pre-setting and extraction of algal features, such as cell size and edge contours. These technologies cannot capture subtle morphological changes in algae under different environments, leading to significant identification errors and an accuracy rate below 50% in complex scenarios.

[0005] Adaptability: Natural water samples contain many impurities and have uneven lighting, which traditional models struggle to handle. Impurities such as sediment, zooplankton, surface reflections, and underwater shadows in the water can interfere with algae identification, causing traditional models to easily misidentify impurities as algae or miss algae altogether, resulting in extremely poor adaptability to complex scenarios.

[0006] In terms of update capabilities: When new algal species appear or existing algal morphological changes occur, traditional basic image recognition technology requires manual re-collection of samples, annotation of features, and full training of the model, which takes several weeks and cannot adapt to the dynamic changes of algae in a timely manner, seriously lagging behind actual monitoring needs.

[0007] Therefore, a smart system and method for recognizing multidimensional features of algae based on a large model is needed to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent recognition system and method for multi-dimensional features of algae based on a large model, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent recognition system for multi-dimensional features of algae based on a large model, comprising:

[0010] The data acquisition module consists of an underwater camera, a microscope imaging system, and environmental sensors. It collects data from multiple dimensions, including macroscopic distribution, microscopic morphology, and environmental correlation, and is used to acquire raw images, microscopic morphology videos, and environmental parameters of algae.

[0011] The preprocessing module, connected to the data acquisition module, is used to perform denoising, enhancement, and segmentation operations on the raw data, and output clear algal cell images, keyframe videos, and standardized environmental parameters.

[0012] The feature extraction module, deployed on the GPU server and connected to the preprocessing module, is used to extract static features (cell size ±0.1μm, shape), dynamic features (motion trajectory ±0.5μm / s), and environmental features from the processed data, and generate standardized feature vectors to solve the problem of single feature dimensions in traditional methods.

[0013] The feature library module, connected to the feature extraction module, uses a distributed database and feature indexing tools to store feature vectors with a retrieval speed of ≤50ms, providing historical data references for inference and improving the ability to identify new samples.

[0014] The large model inference module is connected to the feature extraction module and the feature library module respectively, and adopts an improved convolutional neural network (CNN) + long short-term memory network (LSTM) hybrid structure to infer the feature vector;

[0015] The result verification module is connected to the large model inference module. It processes results according to confidence level. High confidence (≥90%) results are output directly, medium confidence (60%-90%) results are output after comparison with the expert annotation library, and low confidence (<60%) results are pushed to experts for review, thus balancing efficiency and accuracy.

[0016] The incremental training module is connected to the result verification module and the large model inference module respectively. It receives new samples after expert review, freezes the parameters of the first 10 CNN layers and only updates the classification layer. It adopts the SGD optimizer (learning rate 0.001), with a single training session of ≤4 hours and completes model updates within 24 hours, achieving rapid optimization.

[0017] The real-time monitoring terminal module is connected to the result verification module and is used to present the identification results in real time. When toxic algae are detected, the early warning response is ≤3 seconds, providing timely support for monitoring decisions.

[0018] In a further technical solution, the underwater camera and the preprocessing module are connected via USB 3.0 wired connection in the data acquisition module, the microscope imaging system is connected to the preprocessing module via HDMI cable, and the environmental sensor communicates with the preprocessing module via a 4G wireless module; the underwater camera has 5 megapixels, supports 10x optical zoom, and can acquire images at magnifications of 100×, 400×, and 1000×, as well as under bright field, dark field, and fluorescence illumination conditions; the environmental sensor includes a water temperature sensor, a pH sensor, and a dissolved oxygen sensor, with accuracies of ±0.1℃ and ±0.01pH, respectively.

[0019] In a further technical solution, the preprocessing module uses a wavelet transform algorithm with db4 wavelet basis and 3-level decomposition as parameters for noise reduction, employs an adaptive histogram equalization algorithm to enhance image contrast, and segments cell contours using the maximum inter-class variance thresholding method; the preprocessing module is mounted on an edge computing terminal equipped with an Intel Core i7 processor and 8GB of memory.

[0020] In a further technical solution, the CNN in the large model inference module is based on ResNet50 with three additional 1×1 convolutional layers, and the LSTM adopts a two-layer 128 hidden unit structure. The large model inference module is deployed on a high-performance server cluster, each equipped with eight NVIDIA A100 graphics cards, using a distributed inference framework, with an inference latency of ≤1 second / sample and a computing power of ≥10 PFlops.

[0021] A further technical solution is that the incremental training module achieves rapid model optimization by freezing the bottom feature layer (parameters of the first 10 CNN layers) and updating only the classification layer parameters, with a single incremental training time of ≤4 hours; the SGD optimizer is used, the learning rate is set to 0.001, and the model update cycle is ≤24 hours.

[0022] A method for intelligent recognition of multi-dimensional features of algae based on a large model, applied to any of the aforementioned intelligent recognition systems for multi-dimensional features of algae based on a large model, includes the following steps:

[0023] S1. Data Acquisition and Preprocessing Stage: The data acquisition module is used to acquire the original images, videos and environmental parameters. After wavelet transform denoising with db4 wavelet basis and 3-level decomposition as parameters, adaptive histogram equalization and threshold segmentation are performed to output clear images, key frames and standardized parameters. For microscopic morphology videos, the inter-frame difference method is used to extract key frames, extracting 1 frame at 5-frame intervals.

[0024] S2, Feature Extraction and Matching Stage: Extract static features, dynamic features, and environmental-related features from the processed data, generate feature vectors, and retrieve similar historical samples through the feature library module;

[0025] S3, Model Inference and Result Verification Stage: Input the feature vector into the CNN+LSTM hybrid model for inference, process the results according to the confidence threshold, directly output high confidence results, output medium confidence results after comparing with the expert annotation library, and push low confidence results to experts for review.

[0026] S4. Model Dynamic Update Phase: After receiving new samples reviewed by experts, the model's bottom feature layer is frozen, and only the classification layer parameters are updated for incremental training. The model parameters and feature library are updated synchronously. The SGD optimizer is used with a learning rate of 0.001, and the model update cycle is ≤24 hours.

[0027] In a further technical solution, in step S2, the extracted static features include cell size (accuracy ±0.1μm), shape, color channel mean, and cell wall texture; the dynamic features include motion trajectory (velocity accuracy ±0.5μm / s) and cell division time sequence features.

[0028] In a further technical solution, in step S3, an attention mechanism and a contrastive learning algorithm are introduced during model training, with the contrastive learning using the InfoNCE loss function.

[0029] A further technical solution involves using incremental training to achieve rapid model iteration during the model dynamic update phase, simultaneously updating the model and feature library. In step S3, the high confidence level is ≥90%, the medium confidence level is 60%-90%, and the low confidence level is <60%.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] This invention significantly improves efficiency: through a fully automated design (data acquisition-preprocessing-inference-verification-update), combined with edge computing, GPU parallel computing, and high-performance server clusters, the single-sample identification time is reduced from 15 minutes for manual identification to 5 seconds, achieving a 180-fold increase in efficiency; it can process over 100,000 samples per day, meeting the needs of large-scale water body monitoring (such as watershed surveys), and solving the problems of low efficiency and inability to handle large-scale samples in traditional methods;

[0032] This invention optimizes accuracy by employing multi-dimensional feature fusion (static + dynamic + environmental association), an improved hybrid model (CNN + LSTM + attention + contrastive learning), and a hierarchical verification mechanism. This results in an accuracy rate of ≥98% for common algae identification, ≥95% for similar species differentiation, and ≥90% for identification in complex scenarios (including impurities, low light, and algae overlap), thus solving the problems of low accuracy and high misjudgment rate in traditional identification methods.

[0033] This invention features enhanced scene adaptability: environmental interference filtering algorithms (wavelet denoising, adaptive equalization) and targeted processing strategies (such as overlapping algae separation and complex lighting enhancement) enable the system to maintain high recognition accuracy in complex scenarios such as those containing silt and impurities, uneven lighting, and dense overlapping algae. This overcomes the problem of poor scene adaptability of traditional technologies and ensures the effective implementation of monitoring work in different water environments.

[0034] This invention features dynamic update capabilities: an incremental training mechanism and a continuously expanding feature library strategy, which shortens the time for incorporating new algal species from several weeks to 24 hours. Furthermore, incremental training only updates the classification layer parameters, reducing computational resource consumption by 60%. The feature library is continuously expanded through academic literature crawling and manual annotation, ensuring that the system can quickly adapt to dynamic changes in algae (the emergence of new species, morphological variations), maintain long-term recognition capabilities and accuracy, and solve the problems of traditional models being slow to update and unable to adapt to changes in algae.

[0035] This invention features modular synergy: the preprocessing module collaborates with the feature extraction module to improve feature extraction accuracy by 15%, leading to an 8% increase in model inference accuracy; the feature library module works in conjunction with the large model inference module to double the inference speed; and the incremental training module and the large model inference module iteratively optimize to ensure no decay in model accuracy within a one-year period, thereby improving the overall system performance and solving the problems of traditional modules operating independently and having poor synergy.

[0036] This invention offers practical value for monitoring and early warning: the real-time monitoring terminal visually presents algae identification results through a graphical interface (distribution heat map, species percentage pie chart), triggering an audible and visual alert plus SMS warning within 3 seconds when toxic algae are detected, and the data is traceable and analyzable; it provides strong support for water environment monitoring and early warning decision-making, helping users to promptly grasp changes in water quality (such as algae outbreaks and toxic algae invasions), and solving the problems of lagging traditional monitoring and early warning and insufficient decision-making basis. Attached Figure Description

[0037] Figure 1 This is a system flowchart of the present invention;

[0038] Figure 2 This is a diagram illustrating the working steps of the present invention;

[0039] Figure 3 This is a schematic diagram of the data acquisition and preprocessing process of the present invention;

[0040] Figure 4 This is a schematic diagram of the feature extraction and matching process of the present invention;

[0041] Figure 5 This is a schematic diagram of the process for model thrust and verification in this invention;

[0042] Figure 6 This is a schematic diagram illustrating the process of dynamically updating the model of this invention. Detailed Implementation

[0043] The present invention will be further described below with reference to embodiments.

[0044] The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions, and simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0045] Example 1: Watershed-wide water quality survey scenario (taking monthly monitoring of tributaries of the Yangtze River Basin as an example)

[0046] like Figures 1-6 As shown, this embodiment of the invention provides an intelligent algae multi-dimensional feature recognition system based on a large model, including a deployment architecture design, specifically divided into a front-end acquisition layer, an edge processing layer, a cloud inference layer, and a terminal application layer:

[0047] Underwater camera configuration: Select an underwater camera with 10x optical zoom and 5 megapixels (such as the XC-500UW model), install it on a custom bracket 1.5 meters underwater, set it to automatically take a panoramic image once every 2 hours, and connect it directly to the edge computing terminal via USB 3.0 interface with a transmission bandwidth of ≥500Mbps;

[0048] Microscope imaging system: Equipped with a 1000× oil immersion microscope and supporting HDMI lossless output, such as the MC-2000 model, for water samples collected on site, three samples are randomly selected for each monitoring cycle (month) and a 5-minute video recording of the micromorphology of each sample is performed. The video data is transmitted to the edge terminal in real time via HDMI cable.

[0049] Environmental sensor cluster: integrates multiple parameter sensors such as water temperature (accuracy ±0.1℃), pH (accuracy ±0.01), and dissolved oxygen (accuracy ±0.2mg / L), collects water quality data at a frequency of 1 minute / time, and synchronizes it to the edge computing terminal via a 4G wireless module (transmission delay ≤200ms).

[0050] Edge processing layer: One edge computing terminal (configured with an Intel Core i7-12700 processor and 32GB of memory) is deployed at each monitoring section to handle data preprocessing and preliminary feature extraction tasks.

[0051] Preprocessing steps: Wavelet transform algorithm (db4 wavelet basis, 3-level decomposition) is used for noise reduction (noise suppression ratio ≥35dB), adaptive histogram equalization is used to improve contrast (gain ≥3 times), and OTSU threshold segmentation method is used to separate algae from the background (cell contour extraction accuracy ≥98%).

[0052] Lightweight feature extraction: Basic morphological features are extracted based on the pre-trained MobileNetV3 model to generate a preliminary 32-dimensional feature vector.

[0053] Cloud-based inference layer: A high-performance inference cluster is built on the cloud computing center, including two 8-GPU NVIDIA A100 servers (each GPU has a computing power of ≥312 TFLOPS), deploying a complete large-scale algae recognition model.

[0054] Feature library construction: A feature library covering more than 300 algae species in the Yangtze River Basin was built based on the MongoDB distributed database, storing more than 100,000 feature vectors. The FAISS indexing tool was used to achieve a retrieval latency of ≤50ms.

[0055] Large-scale model inference: Deploy an improved CNN+LSTM hybrid model (ResNet50 + 3 layers of 1×1 convolution to extract static features, and two layers of LSTM to analyze dynamic temporal sequences), integrating attention mechanism and contrastive learning algorithm (InfoNCE loss function), achieving single-sample inference latency ≤1 second and common algae identification accuracy ≥98%;

[0056] Incremental training mechanism: Low-confidence samples with expert review and annotation are pulled every week, the parameters of the first 10 CNN layers of the model are frozen, and the model is iterated and updated within 4 hours using the SGD optimizer (learning rate 0.001).

[0057] Terminal application layer:

[0058] The monitoring center is visualized by using a 55-inch LED splicing screen to construct a "heat map of algae distribution in the watershed", a "dynamic pie chart of species proportion", and a "real-time early warning bar for toxic algae". When the proportion of toxic algae is ≥5%, an audible and visual alarm and SMS push are triggered (response time ≤3 seconds).

[0059] Data traceability and analysis: Build a web-based platform that supports querying historical data and feature database update records by time, river section, and algae species.

[0060] In this embodiment, the system achieves a single water sample processing time of ≤10 minutes, automatic identification of over 100,000 samples from 10 cross-sections per month, and a reduction in labor costs of 80%. The accuracy rate for identifying common algae reaches 98.2%, the accuracy rate for distinguishing easily confused species is 95.5%, and the accuracy rate in complex scenarios is 90.3%. The system updates the model within 24 hours for newly invasive algae, achieving an accuracy rate of ≥95%. Over 500,000 algae samples have been identified cumulatively, monthly distribution maps of algae in the Yangtze River tributaries have been created, and early warnings for cyanobacterial blooms are provided two days in advance, helping to reduce the impact area by 70%.

[0061] Example 2: Algae monitoring scenario in aquaculture ponds (taking daily prevention and control of 10,000 mu shrimp ponds as an example)

[0062] The difference between this embodiment and Embodiment 1 is that it adopts a lightweight deployment architecture, simplifies the data acquisition terminal and inference process, and adapts to the fragmented monitoring needs and low-cost operation and maintenance scenarios of aquaculture ponds. This embodiment of the invention provides an intelligent algae multi-dimensional feature recognition system based on a large model. The system includes a lightweight deployment scheme, specifically divided into a pond data acquisition terminal and a cloud-based simplified inference module.

[0063] Pond monitoring terminals: One monitoring terminal is deployed at every 50 mu (approximately 3.3 hectares) of aquaculture ponds, integrating:

[0064] Portable microscopic imaging unit: It adopts a handheld microscopic device (such as PM-1000 model) with a 400× objective lens and supports wireless transmission. After the aquaculture farmers collect water samples from the pond, they take microscopic morphological photos / videos and transmit them directly to the cloud via Wi-Fi.

[0065] Miniature environmental sensor: integrates water temperature, pH, and dissolved oxygen sensors (size ≤ fist size), automatically collects water quality data once every 30 minutes, and transmits it back to the pond edge gateway via LoRa wireless module (transmission distance ≥ 2km).

[0066] Cloud-based simplified reasoning:

[0067] Customized feature library: A lightweight feature library (storing ≤20,000 feature vectors) was built for 200+ common algae in aquaculture ponds, and the inference model parameters were optimized (redundant network layers were pruned) to achieve a single-sample inference latency of ≤0.5 seconds;

[0068] Early warning rule configuration: Based on aquaculture experience, early warning thresholds such as "green algae percentage <30%" and "dinoflagellate percentage >3%" are set. When the monitoring data triggers the threshold, operation guidance and early warning information (such as "pond algae imbalance, it is recommended to turn on the aerator") are pushed through the pond LED screen and the farmer's mobile APP.

[0069] In this embodiment, the system achieves a total process time of ≤5 minutes from algal anomaly identification to early warning reception, with an accuracy rate of ≥92% for identifying water samples from complex ponds. In a case of "algal collapse" early warning in a shrimp pond, the risk was identified 3 days in advance, reducing economic losses by over 200,000 yuan; replacing the traditional manual microscopic examination method, farmers save 2 hours per day, and the yield per mu of ponds increases by 12%; by establishing a "algal change-disease occurrence" correlation database, the morphological changes of dinoflagellates can be traced back to 3 days before the abnormality in diseased ponds, verifying the value of early disease warning and promoting the transformation of aquaculture models from "passive disease treatment" to "proactive prevention and control".

[0070] Working principle and usage process of this invention:

[0071] By constructing a multi-source data fusion preprocessing mechanism, a multi-dimensional feature extraction and hybrid model inference architecture, a confidence level verification and model dynamic update system, intelligent identification of multi-dimensional features of algae is achieved, specifically including the following steps:

[0072] First, data collaboration and optimization are achieved through multi-source data fusion and preprocessing mechanisms: At the data acquisition layer, underwater cameras construct an acquisition network according to the dimensions of "macro-micro-environment." The underwater cameras periodically capture the aggregation morphology and distribution range of algae in different areas of the water body at a resolution of ≥5 megapixels. The microscope imaging system captures algal cell structure and movement trajectories with a 1000× high-power objective lens and a frame rate of ≥25fps, generating microscopic dynamic videos. Environmental sensors collect parameters such as water temperature, pH, and dissolved oxygen at high frequencies to establish a "water quality conditions-algae" model. The data is associated with the morphology of the algae. A preprocessing algorithm chain is executed sequentially on the collected multi-source data. Denoising is performed using db4 wavelet basis 3-layer decomposition to improve the signal-to-noise ratio to ≥45dB. Adaptive histogram equalization is used to improve the contrast gain to ≥3 times for underwater low light and reflective scenes to highlight the edge and internal structure of algae. Based on the OTSU threshold segmentation method, algal cells and background are segmented with an accuracy of ≥98% and a standardized image of 1024×1024 pixels is output. At the same time, the environmental parameters are normalized to construct a "water quality-algae" feature association vector.

[0073] Secondly, feature parsing and accurate inference are achieved through feature extraction and hybrid model inference logic: In multi-dimensional feature parsing, the feature extraction module, based on the computing power of the GPU server, constructs a three-dimensional feature system of "static + dynamic + environment" from the preprocessed data, extracting cell size (accuracy ±0.1μm), shape, color channel mean, cell wall texture, etc. to generate a 256-dimensional static feature vector. A 128-dimensional dynamic feature vector is constructed by tracking the motion trajectory using optical flow (velocity accuracy ±0.5μm / s) and analyzing the cell division cycle (division interval error ≤10 seconds). Finally, environmental parameters are fused to establish a mapping relationship and generate a 128-dimensional ring. The environmental feature vectors are concatenated to form a 512-dimensional standardized feature vector. In the hybrid model inference architecture, the large model inference module adopts an improved CNN+LSTM hybrid structure. The CNN branch uses ResNet50 as the base network with 3 layers of 1×1 convolutional layers. The attention mechanism assigns high weights to the special pigment regions of toxic algae to enhance static feature extraction. The LSTM branch designs a two-layer LSTM network (128 hidden units per layer) to perform temporal modeling of keyframe sequences in microscopic video to analyze dynamic features. At the same time, the InfoNCE loss function is introduced to expand the inter-class feature distance, thereby improving the inference accuracy to ≥98%.

[0074] Furthermore, result optimization and system evolution are achieved through confidence-level verification and dynamic model updates: In the three-level confidence verification process, results with ≥90% high confidence are directly pushed to the real-time monitoring terminal due to their high similarity to historical samples in the feature library and environmental parameter matching degree ≥95%; results with 60%-90% medium confidence are triggered by environmental interference, triggering a "secondary search of the feature library," and are pushed after comparison with the expert annotation library to correct misjudgments; results with <60% low confidence are automatically triggered by the lack of similar samples, triggering an "expert review process," where experts review and compare with literature to label the correct categories and supplement features to form new samples; In incremental training and model iteration, after receiving new samples, the parameters of the first 10 CNN layers of the model are frozen, and only the classification layer is updated. The SGD optimizer (learning rate 0.001) is used to complete the iteration within 4 hours. At the same time, ≥5 new feature entries are added weekly according to the "morphology + environment" dimension to expand the feature library. After the update, the accuracy of new algae recognition is ≥95%, the accuracy of complex scene recognition is ≥90%, and the inference latency is ≤1 second / sample through test set verification;

[0075] Finally, by fusing multi-dimensional features, the limitations of traditional single image recognition are overcome. By combining microscopic dynamics and environmental parameters, the problem of algae identification is solved. Through a closed loop of "low confidence samples - expert annotation - incremental training", the model can adapt to new algae and new scenarios within 24 hours. Through lightweight deployment and customized feature library, it meets the needs of different scenarios such as watershed water quality monitoring and aquaculture prevention and control, providing intelligent tools for water environment management and aquaculture efficiency improvement.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various 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 multi-dimensional feature intelligent recognition system for algae based on a large model, characterized in that, include: The data acquisition module consists of an underwater camera, a microscope imaging system, and environmental sensors, and is used to acquire raw images, microscopic morphological videos, and environmental parameters of algae. The preprocessing module, connected to the data acquisition module, performs denoising, enhancement, and segmentation operations on the raw data, outputting clear algal cell images, keyframe videos, and standardized environmental parameters. The feature extraction module, connected to the preprocessing module, is used to extract static features, dynamic features, and environmental features from the processed data and generate standardized feature vectors. The feature library module, connected to the feature extraction module, is used to store feature vectors and support similar feature sample queries; The large model inference module is connected to the feature extraction module and the feature library module respectively. It adopts an improved convolutional neural network (CNN) + long short-term memory network (LSTM) hybrid structure to infer the feature vector and output the algae species determination result and confidence level. The result verification module is connected to the large model inference module and performs hierarchical processing on the inference results based on the confidence threshold. The incremental training module is connected to the result verification module and the large model inference module respectively, and receives new samples after expert review to achieve rapid model optimization. The real-time monitoring terminal module is connected to the result verification module and is used to present the identification results in real time and trigger an early warning when toxic algae are detected.

2. The intelligent recognition system for multi-dimensional features of algae based on a large model according to claim 1, characterized in that, In the data acquisition module, the underwater camera and the preprocessing module are connected via USB 3.0 wired connection, the microscope imaging system is connected to the preprocessing module via HDMI cable, and the environmental sensor communicates with the preprocessing module via a 4G wireless module. The underwater camera has 5 megapixels, supports 10x optical zoom, and can acquire images at magnifications of 100×, 400×, and 1000×, as well as under bright field, dark field, and fluorescence illumination conditions. The environmental sensor includes a water temperature sensor, a pH sensor, and a dissolved oxygen sensor, with accuracies of ±0.1℃ and ±0.01pH, respectively.

3. The intelligent recognition system for multi-dimensional features of algae based on a large model according to claim 1, characterized in that, The preprocessing module uses a wavelet transform algorithm with db4 wavelet basis and 3-level decomposition as parameters for noise reduction, employs an adaptive histogram equalization algorithm to enhance image contrast, and segments cell contours using the maximum inter-class variance thresholding method. The preprocessing module is mounted on an edge computing terminal equipped with an Intel Core i7 processor and 8GB of memory.

4. The intelligent recognition system for multi-dimensional features of algae based on a large model according to claim 1, characterized in that, The large model inference module uses ResNet50 as the base network for CNN and adds three 1×1 convolutional layers. The LSTM adopts a two-layer 128 hidden unit structure. The large model inference module is deployed on a high-performance server cluster, each equipped with eight NVIDIA A100 graphics cards. It uses a distributed inference framework, with an inference latency of ≤1 second / sample and a computing power of ≥10PFlops.

5. The intelligent recognition system for multi-dimensional features of algae based on a large model according to claim 1, characterized in that, The incremental training module achieves rapid model optimization by freezing the bottom feature layer (parameters of the first 10 CNN layers) and updating only the classification layer parameters, with a single incremental training time of ≤4 hours; it uses the SGD optimizer with a learning rate set to 0.001, and the model update cycle is ≤24 hours.

6. A method for intelligent recognition of multi-dimensional features of algae based on a large model, applied to the intelligent recognition system for multi-dimensional features of algae based on a large model as described in any one of claims 1-5, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing Stage: The data acquisition module is used to acquire the original images, videos and environmental parameters. After wavelet transform denoising with db4 wavelet basis and 3-level decomposition as parameters, adaptive histogram equalization and threshold segmentation are performed to output clear images, key frames and standardized parameters. For microscopic morphology videos, the inter-frame difference method is used to extract key frames, extracting 1 frame at 5-frame intervals. S2, Feature Extraction and Matching Stage: Extract static features, dynamic features, and environmental-related features from the processed data, generate feature vectors, and retrieve similar historical samples through the feature library module; S3, Model Inference and Result Verification Stage: The feature vectors are input into the CNN+LSTM hybrid model for inference. The results are processed in stages according to the confidence threshold. High-confidence results are directly output, medium-confidence results are compared with the expert annotation library and then output, and low-confidence results are sent to experts for review. S4. Model Dynamic Update Phase: After receiving new samples reviewed by experts, the model's bottom feature layer is frozen, and only the classification layer parameters are updated for incremental training. The model parameters and feature library are updated synchronously. The SGD optimizer is used with a learning rate of 0.001, and the model update cycle is ≤24 hours.

7. The intelligent recognition method for multi-dimensional features of algae based on a large model according to claim 6, characterized in that, In step S2, the extracted static features include cell size (accuracy ±0.1μm), shape, color channel mean, and cell wall texture; the dynamic features include motion trajectory (velocity accuracy ±0.5μm / s) and cell division time sequence features.

8. The intelligent recognition method for multi-dimensional features of algae based on a large model according to claim 6, characterized in that, In step S3, an attention mechanism and a contrastive learning algorithm are introduced during model training. The contrastive learning uses the InfoNCE loss function.

9. The intelligent recognition method for multi-dimensional features of algae based on a large model according to claim 6, characterized in that, During the model dynamic update phase, incremental training is used to achieve rapid model iteration and synchronously update the model and feature library. In step S3, the high confidence level is ≥90%, the medium confidence level is 60%-90%, and the low confidence level is <60%.

Images