Method for detecting and identifying two insects in water based on deep learning

Abstract

The invention discloses a method for detecting and identifying two insects in water based on deep learning, and relates to the technical field of aquatic biological detection.The method comprises the steps that insect bodies are enriched through a micro-fluidic chip, DAPI / PI dyeing is automatically completed, and recognizable biochemical fluorescence characteristics are given to the insect bodies; dAPI, PI and bright field multichannel images are acquired through a CMOS camera, and sub-pixel-level alignment of multispectral images is realized by using an image registration and fusion algorithm; denoising and segmenting the multispectral image, extracting an insect body area and enhancing the contrast; the positions of polypide bodies are detected through an improved YOLOv8 model, and the activity states of the polypide bodies are judged in combination with fluorescence characteristics and form scores; counting the number of insect bodies and the live-to-dead ratio in real time, identifying capsules, and triggering live insect standard exceeding alarm and pollution trend analysis; a model is deployed on an edge device, an image uploading result is compressed, and data security and traceability are ensured through a block chain. The system integrates sample processing, imaging acquisition, intelligent identification and data feedback, and is suitable for long-time stable operation on site.

Description

Technical Field

[0001] The present invention relates to the technical field of aquatic organism detection, and in particular to a method for detecting and identifying two insects in water based on deep learning. Background Art

[0002] Cryptosporidium and Giardia lamblia are two pathogenic protozoa that parasitize humans and animals. They are highly resistant to chlorine disinfection in drinking water and have repeatedly caused large-scale waterborne disease outbreaks. These two parasites pose a direct threat to drinking water safety and have garnered widespread attention worldwide. my country's revised "Standards for Drinking Water Quality" now include Cryptosporidium and Giardia lamblia in drinking water microbial indicators and strictly limits their levels, demonstrating the importance of monitoring these two parasites in water for health.

[0003] Currently, the detection of Cryptosporidium and Giardia lamblia in water primarily follows the US Environmental Protection Agency's EPA 1623 standard method. This method involves filtration and enrichment of large-volume water samples, isolation and purification of the parasites, fluorescent antibody staining, and microscopic examination. In the traditional testing process, the water sample is typically filtered and concentrated using a membrane or filter. Immunomagnetic bead technology is then used to separate Cryptosporidium oocysts and Giardia lamblia cysts from the concentrate. Fluorescently labeled specific antibodies are then used to stain the parasites, and finally, the parasites are observed, identified, and counted using a fluorescence microscope. This method offers high sensitivity and accuracy, but also has significant limitations.

[0004] First, the traditional "two insects" detection process is complex and time-consuming, requiring special filter materials, reagents and microscopic imaging equipment. The detection cost is high and relies on imported instruments and consumables. Secondly, the interpretation of the results is highly dependent on manual labor: professional technicians need to identify the worms based on morphological characteristics under a microscope. This process is highly subjective and easily affected by human experience. In addition, there is a shortage of professionals with this skill in China, which restricts the implementation of large-scale monitoring. Thirdly, existing methods are difficult to provide test results in a timely manner. It usually takes several hours to several days from sampling to obtaining a laboratory test report, which cannot meet the needs of real-time early warning on site. In addition, it is also difficult to judge the vitality of the worms. Existing technologies usually require the use of staining methods to indirectly infer whether the worms are alive, but these steps also require manual analysis, and real-time determination of parasite activity is challenging.

[0005] To address the above problems, the present invention proposes a method for detecting and identifying two insects in water based on deep learning. Summary of the Invention

[0006] In response to the above problems, the present invention provides a method for detecting and identifying two insects in water based on deep learning to solve the problems of low efficiency, reliance on manual labor, and inability to identify activity in real time in the existing technology of "two insects" detection.

[0007] To solve the above technical problems, the present invention provides the following technical solution: a method for detecting and identifying two insects in water based on deep learning, comprising the following steps: Step S1, enriching the parasites in the water sample and removing impurities using a microfluidic chip, and using DAPI and PI dyes to label the biochemical status of the parasites; In step S1, the following sub-steps are also included: S1-1: The water sample to be tested is introduced into a microfluidic chip with micron-scale channels. The sample is filtered through two filter membranes: the first filter with a 50μm pore size removes large algae particles and impurities, while the second filter with a 5μm pore size retains target parasites of 3-6μm. At a flow rate of 100-200μL / min, inertial fluid dynamics is used to force the parasites into a monolayer in the central channel. S1-2, freeze-dried DAPI / PI tablets were injected into PBS solution via a piezoelectric micropump to form a stable concentration, incubated at a constant temperature of 37±0.2°C for 20 min, and three-stage PBS washing was used to eliminate dye residues.

[0008] Step S2, by collecting DAPI, PI and bright field multi-channel images, using image registration and fusion algorithms to achieve sub-pixel alignment of the multi-channel images; In step S2, the following sub-steps are also included: S2-1, equipped with a 6-band LED array and an electrically controlled filter wheel, uses PWM to modulate the light source intensity to prevent fluorescence quenching. The CMOS camera uses adaptive exposure, automatically adjusting according to the initial frame signal to prevent low-fluorescence insects from being overlooked; S2-2, based on the improved SIFT algorithm, performs sub-pixel alignment of multispectral images with a registration error of less than 0.3 pixel, and outputs a 4-channel composite image, specifically bright field, DAPI, PI, and reflected light; S2-3, calculate the ratio of the fluorescence intensity variance to the background intensity of each frame image. If it is less than 15dB, trigger automatic delayed exposure; if the insect body confidence in the detection results is less than 0.3, it is judged as imaging abnormality and re-staining is performed.

[0009] Step S3: Data cleaning and optimization are performed through adaptive image enhancement and background noise segmentation, dynamic selection of denoising algorithms using image histograms, enhancement of DAPI / PI contrast, and mask generation and ROI extraction based on a lightweight UNet structure. In step S3, the following sub-steps are also included: S3-1, fusing the multispectral image data obtained in step S2, overlapping the insect body shape provided by the bright field image with the internal structure and highlighted areas reflected by the fluorescent image, to form composite image data containing richer information; Image histogram statistics are used to infer the current image contrast and noise level, and a noise reduction scheme is dynamically selected. An improved Retinex algorithm is used to enhance the local contrast of the DAPI / PI channel while preserving the edge information of the insect body to avoid artifacts caused by over-sharpening. S3-2, based on the Mobile-UNet structure, realizes the binary mask generation of algae, bubbles, and water stains; generates the minimum bounding rectangle based on the segmentation results to limit the detection area.

[0010] Step S4: Input the fused image into the improved YOLOv8 deep learning model, combine it with the CBAM attention mechanism to achieve multi-scale insect body localization, and improve the frame fit through rotation frame regression; In step S4, the following sub-steps are also included: S4-1, based on the improved YOLOv8 model, embeds the CBAM attention mechanism in the backbone network, uses rotating frame regression, adds a small target detection head, supports detecting more than 30 insect bodies in a single frame, and regresses the rotation angle of the detection frame according to the main axis direction of the insect body to avoid background fluorescence interference; S4-2, quantitative analysis of fluorescence signals, by extracting the mean PI channel intensity, DAPI nuclear-cytoplasmic ratio, and fluorescence area ratio, setting a decision threshold to assist in determining vitality; quantitative scoring of edge discontinuity, central voids, and contour collapse features to determine death status, and using temporal ambiguity analysis to determine live worms for low-confidence samples; S4-3, through the uncertainty processing mechanism, multiple rounds of reasoning calculation output confidence distribution. Confidence variance greater than 0.1 is considered a high uncertainty sample, not included in the statistical analysis, and marked as "manual review required".

[0011] Step S5: Real-time statistics and time series analysis of the test results are used to count the number of insects and the live-to-dead ratio in real time, identify cysts, and trigger an alarm for excessive live insects and conduct pollution trend analysis. In step S5, the following sub-steps are also included: S5-1: Update the live / dead count and ratio of parasites every 10 seconds, generate a 10-minute sliding window trend curve, and detect mutation points in vitality; set up a cyst identification mechanism: weak DAPI positive, PI negative, rounded edges, and no internal structure are considered cysts and classified as "potential transmission risk"; S5-2, concentration warning mechanism: if the live insect concentration is greater than 1 / 10L, the system will automatically issue a red alarm and upload the information to the supervision system API; Conduct event trend analysis, use the relationship between live insect concentration and time, and use the exponential model to determine whether the pollution is sudden or continuous.

[0012] In step S6, the model is deployed to an edge computing device that supports TensorRT acceleration, and efficient inference is achieved by combining FP16 quantization and dynamic resolution adjustment.

[0013] In step S6, the following sub-steps are also included: S6-1 deploys the TensorRT inference engine, supports FP16 quantization and Layer Fusion, and increases inference speed to over 30FPS; it also supports dynamic input switching to adapt to different computing power; S6-2 crops the detection area and compresses it using JPEG 2000, transmitting only the valid area, reducing bandwidth usage by more than 50%. It also supports LoRaWAN remote upload, uploading insect count, activity ratio, and warning status in a package every hour. S6-3, the test results are uploaded to the Ethereum private chain through SHA-256 hash encryption to ensure data security and traceability, and the number of times the sample passes is recorded. If it exceeds 50 times, it will automatically prompt cleaning to avoid cross contamination caused by residues.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses a microfluidic chip to achieve automatic enrichment and pretreatment of water samples, combined with fully automatic fluorescence staining and multispectral imaging processes, which greatly shortens the detection cycle and achieves rapid sample processing and image acquisition.

[0015] The present invention significantly improves the detection accuracy and anti-interference ability by integrating the biochemical characteristics of DAPI / PI staining with the morphological information of bright field images, and combining it with a deep learning model to accurately identify the insect position, type and vitality status.

[0016] While detecting the insect body, the present invention combines fluorescence intensity quantification and morphological analysis to automatically distinguish between live insects, dead insects and cyst states, and achieve objective judgment of insect body activity, breaking through the limitation of traditional methods that rely on manual interpretation.

[0017] Through deep learning models and dynamic activity analysis deployed on edge devices, the present invention can run continuously on site and issue real-time warnings of excessive insect concentrations, supporting pollution trend assessment and emergency response.

[0018] The test results of the present invention are uploaded in a compressed and encrypted manner, and are stored on the chain in combination with blockchain technology to ensure that the data cannot be tampered with and have good traceability and regulatory compliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It is understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but is merely for selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] Please refer to Figure 1 This is a schematic diagram of a method for detecting and identifying two insects in water based on deep learning provided by an embodiment of the present invention, including: Step S1: Enrich the parasites in the water sample and remove impurities using a microfluidic chip, and use DAPI and PI dyes to mark the biochemical status of the parasites.

[0023] S1-1, introduce the water sample to be tested into the microfluidic chip pretreatment module. The microfluidic chip is integrated with a micron-level channel structure and a filter element to intercept and enrich the target insects in the water sample.

[0024] A double-layer membrane structure is set up, with the first-level pore size of 50μm intercepting large particles of impurities such as algae and fibers; the second-level pore size of 5μm fine screening retains insect egg cysts with a diameter of 3~6μm and effectively removes suspended sediment; using the principle of inertial fluid dynamics, the insect bodies are focused in the central channel at a flow rate of 100~200μL / min, forming a sparse single-layer distribution to avoid overlapping interference in subsequent imaging.

[0025] Specifically, as the water sample flows through the chip channels, larger impurities are filtered out, while tiny parasites such as Cryptosporidium oocysts and Giardia lamblia cysts are trapped and concentrated in specific areas, thereby improving the sensitivity of subsequent detection. Compared to traditional large-volume membrane filtration, microfluidic enrichment offers the advantages of low sample consumption and a high degree of automation, allowing for continuous on-site water sample processing.

[0026] Sampling volume and time are controlled, and the enrichment time for 1L sample is about 10 minutes. Continuous sampling and multi-channel parallel processing are supported. After pretreatment, a sufficient number of target insects have accumulated on the droplets or filter membrane in the insect enrichment area, ready for the next step.

[0027] In step S1-2, the target insects were automatically fluorescently stained within the enrichment area of the microfluidic chip. Lyophilized DAPI / PI tablets were injected into the PBS solution via a piezoelectric micropump to form a stable concentration of 2.0 μg / mL DAPI and 1.0 μg / mL PI. The mixing uniformity was monitored in real time using a microfluidic perturbation module.

[0028] The staining reaction was incubated at a constant temperature of 37±0.2°C for 20 minutes to ensure that DAPI fully penetrated the cell nucleus and bound to DNA (mainly labeling the parasite nucleic acid), while at the same time utilizing the integrity of the living cell membrane to prevent PI penetration (only dead parasites were labeled by PI); unbound dye was removed by rinsing with PBS to achieve differential fluorescent labeling of live / dead parasites.

[0029] A three-stage PBS flushing process of "forward flush + main flush + post-flush" is used, specifically forward flush 5s / main flush 15s / post-flush 5s, to avoid residual fluorescent agent causing signal drift or false positives.

[0030] Step S2, by collecting DAPI, PI and bright field multi-channel images, using image registration and fusion algorithms to achieve sub-pixel alignment of the multi-channel images; S2-1, excitation light source control, is equipped with a 6-band adjustable LED array, which dynamically adjusts the light source intensity through PWM to avoid overexposure and extend the fluorescence life of the dye; the filter wheel uses electronic switching to ensure that the DAPI and PI channel filter switching delay is less than 50ms, and the image acquisition interval is controlled within 100ms.

[0031] Image acquisition was performed using a CMOS camera with an adaptive exposure range of 10–200 ms, which was automatically adjusted based on the initial frame signal to avoid overlooking low-fluorescence insects.

[0032] S2-2, pixel-level alignment, matches key points based on the SIFT algorithm, uses affine transformation to eliminate imaging displacement errors, and achieves sub-pixel alignment of DAPI, PI, and bright field images.

[0033] Multi-channel image structure, the output format is 4-channel image, providing high-dimensional feature space for deep models.

[0034] Bright field images of the worm and particles are acquired under visible light to show the morphology and size of the worm; at the same time, fluorescence signal images are captured under excitation at a specific wavelength to show the position of the stained worm and the fluorescent labeling of its internal structure.

[0035] For different fluorescent dyes, such as the green channel used to mark the insect body and the blue channel indicating vitality, different bands are excited and photographed in sequence to obtain corresponding multispectral images. A series of multispectral images can comprehensively reflect the morphological characteristics and fluorescence characteristics of the insect body in the sample.

[0036] S2-3 calculates the ratio of the fluorescence intensity variance to the background intensity for each frame. If it is less than 15dB, trigger automatic delayed exposure or return to S1 for re-staining. If the confidence level of the insect body in the detection results is less than 0.3, it is judged as an imaging abnormality, and re-staining or adjustment of imaging parameters are required.

[0037] In step S3, the data is cleaned and optimized through adaptive image enhancement and background noise segmentation, and the image histogram is used to dynamically select the denoising algorithm, enhance the DAPI / PI contrast, and perform mask generation and ROI extraction based on a lightweight UNet structure.

[0038] S3-1, inputting the multispectral image data obtained in step S2 into an image processing unit for fusion processing, wherein the morphological outline of the insect body and the fluorescent marker information carried by it are simultaneously presented in a comprehensive image through image fusion, and the insect body shape provided by the bright field image is overlapped with the internal structure and highlighted marked area reflected by the fluorescent image, and the contrast is enhanced, thereby forming composite image data containing richer information. The fusion process can be achieved by using techniques such as image registration, pseudo-color synthesis, or feature stacking; Image histogram statistics are used to infer the current image contrast and noise level, and NLM or wavelet denoising schemes are dynamically selected. An improved Retinex algorithm is used to enhance the local contrast of the DAPI / PI channels while preserving the edge information of the insect body to avoid artifacts caused by over-sharpening.

[0039] S3-2, impurity and background segmentation, based on the Mobile-UNet structure, takes less than 5ms to generate binary masks for algae, bubbles, and water stains. It also generates a minimum bounding rectangle based on the segmentation results to limit the detection area, improving computational efficiency by over 30%.

[0040] It should be noted that multimodal fusion helps improve the accuracy of subsequent algorithms in identifying insects, because deep learning models can simultaneously use morphological and biochemical features to determine the type and vitality of insects. The fusion process can be achieved using techniques such as image registration, pseudo-color synthesis, or feature stacking. After this step, the original multi-channel image is converted into a unified data format suitable for algorithm input, which contains a comprehensive feature representation of each suspected insect target.

[0041] In step S4, the fused image is input into the improved YOLOv8 deep learning model, and the CBAM attention mechanism is combined to realize multi-scale insect body positioning, and the frame fit is improved through rotation frame regression.

[0042] S4-1 is based on an improved YOLOv8 model. The backbone network embeds the CBAM attention mechanism, uses rotating frame regression, and adds a small target detection head. After training on 100,000 images, it has an enhanced response to weakly fluorescent insects and supports detecting more than 30 insects in a single frame. The detection frame regresses the rotation angle according to the main axis direction of the insect body to improve the fitting accuracy and avoid background fluorescence interference.

[0043] S4-2, quantitative analysis of fluorescence signals, by extracting the mean PI channel intensity, DAPI nuclear-cytoplasmic ratio, and fluorescence area ratio, and setting a decision threshold to assist in activity determination.

[0044] Morphological analysis was performed and features such as edge discontinuity, central void, and contour collapse were quantitatively scored to determine the death status; for low-confidence samples, temporal fuzzy analysis was used to determine whether the worms were alive.

[0045] S4-3, through the uncertainty processing mechanism, multiple rounds of reasoning calculation output confidence distribution. Confidence variance greater than 0.1 is considered a high uncertainty sample, not included in the statistical analysis, and marked as "manual review required".

[0046] Step S5: Through real-time statistics and time series analysis of the detection results, the number of insects and the live-to-dead ratio are counted in real time, cysts are identified, and an alarm for excessive live insects and pollution trend analysis are triggered.

[0047] S5-1, real-time statistics and trend backtracking, updating the live / dead count and ratio of the worms every 10 seconds, generating a 10-minute sliding window trend curve, and detecting vitality mutation points; cyst identification mechanism, DAPI weakly positive, PI negative, rounded edges, and no internal structure are considered cysts and classified as "potential transmission risk".

[0048] S5-2, concentration warning mechanism: if the live insect concentration is greater than 1 / 10L, the system will automatically issue a red alarm and upload the information to the regulatory system API; in addition, event trend analysis is conducted, using the relationship between live insect concentration and time to determine whether the pollution is sudden or continuous based on linear regression or exponential models.

[0049] In step S6, the model is deployed to an edge computing device that supports TensorRT acceleration, and efficient inference is achieved by combining FP16 quantization and dynamic resolution adjustment.

[0050] S6-1 deploys the TensorRT inference engine, supports FP16 quantization and LayerFusion, and increases the inference speed to over 30FPS; it also supports dynamic input switching to adapt to different computing power.

[0051] S6-2 crops the detection area and compresses it with JPEG 2000 (compression rate greater than 80%), transmitting only the valid area, reducing bandwidth usage by more than 50%; and supports LoRaWAN remote upload, uploading the number of insects, activity ratio, and warning status in a package once an hour.

[0052] S6-3, the test results are uploaded to the Ethereum private chain through SHA-256 hash encryption to ensure data security and traceability, and the number of times the sample passes is recorded. If it exceeds 50 times, it will automatically prompt cleaning to avoid cross contamination caused by residues.

[0053] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for detecting and identifying two insects in water based on deep learning, characterized in that: The following steps are involved: Step S1, enriching the parasites in the water sample and removing impurities using a microfluidic chip, and using DAPI and PI dyes to label the biochemical status of the parasites; Step S2, by collecting DAPI, PI and bright field multi-channel images, using image registration and fusion algorithms to achieve sub-pixel alignment of the multi-channel images; Step S3: Data cleaning and optimization are performed through adaptive image enhancement and background noise segmentation, dynamic selection of denoising algorithms using image histograms, enhancement of DAPI / PI contrast, and mask generation and ROI extraction based on a lightweight UNet structure. Step S4: Input the fused image into the improved YOLOv8 deep learning model, combine it with the CBAM attention mechanism to achieve multi-scale insect body localization, and improve the frame fit through rotation frame regression; Step S5: Real-time statistics and time series analysis of the test results are used to count the number of insects and the live-to-dead ratio in real time, identify cysts, and trigger an alarm for excessive live insects and conduct pollution trend analysis. In step S6, the model is deployed to an edge computing device that supports TensorRT acceleration, and efficient inference is achieved by combining FP16 quantization and dynamic resolution adjustment.

2. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S1, the following sub-steps are also included: S1-1: The water sample to be tested is introduced into a microfluidic chip with micron-scale channels. The sample is filtered through two filter membranes: the first filter with a 50μm pore size removes large algae particles and impurities, while the second filter with a 5μm pore size retains target parasites of 3-6μm. At a flow rate of 100-200μL / min, inertial fluid dynamics is used to force the parasites into a monolayer in the central channel. S1-2, freeze-dried DAPI / PI tablets were injected into PBS solution via a piezoelectric micropump to form a stable concentration, incubated at a constant temperature of 37±0.2°C for 20 min, and three-stage PBS washing was used to eliminate dye residues.

3. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S2, the following sub-steps are also included: S2-1, equipped with a 6-band LED array and an electrically controlled filter wheel, uses PWM to modulate the light source intensity to prevent fluorescence quenching. The CMOS camera uses adaptive exposure, automatically adjusting according to the initial frame signal to prevent low-fluorescence insects from being overlooked; S2-2, based on the improved SIFT algorithm, performs sub-pixel alignment of multispectral images with a registration error of less than 0.3 pixel, and outputs a 4-channel composite image, specifically bright field, DAPI, PI, and reflected light; S2-3, calculate the ratio of the fluorescence intensity variance to the background intensity of each frame image. If it is less than 15dB, trigger automatic delayed exposure; if the insect body confidence in the detection results is less than 0.3, it is judged as imaging abnormality and re-staining is performed.

4. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S3, the following sub-steps are also included: S3-1, fusing the multispectral image data obtained in step S2, overlapping the insect body shape provided by the bright field image with the internal structure and highlighted areas reflected by the fluorescent image, to form composite image data containing richer information; Use image histogram statistics to infer the current image contrast and noise level, and dynamically select a noise reduction scheme; Use the improved Retinex algorithm to enhance the local contrast of the DAPI / PI channel while preserving the edge information of the insect body to avoid artifacts caused by over-sharpening; S3-2, based on the Mobile-UNet structure, realizes the binary mask generation of algae, bubbles, and water stains; generates the minimum bounding rectangle based on the segmentation results to limit the detection area.

5. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S4, the following sub-steps are also included: S4-1, based on the improved YOLOv8 model, embeds the CBAM attention mechanism in the backbone network, uses rotating frame regression, adds a small target detection head, supports detecting more than 30 insect bodies in a single frame, and regresses the rotation angle of the detection frame according to the main axis direction of the insect body to avoid background fluorescence interference; S4-2, quantitative analysis of fluorescence signals, by extracting the mean PI channel intensity, DAPI nuclear-cytoplasmic ratio, and fluorescence area ratio, and setting a decision threshold to assist in viability determination; Edge discontinuity, central voids, and contour collapse are quantitatively scored to determine the state of death. For low-confidence samples, temporal fuzzy analysis is used to determine whether the worm is alive. S4-3 uses the uncertainty handling mechanism to calculate the output confidence distribution through multiple rounds of reasoning. Confidence variance greater than 0.1 is considered a high-uncertainty sample and is not included in the statistical analysis. It is marked as "requires manual review." 6. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S5, the following sub-steps are also included: S5-1 updates the live / dead count and ratio of parasites every 10 seconds, generates a 10-minute sliding window trend curve, and detects sudden changes in activity. A cyst identification mechanism is established: weak DAPI positivity, PI negativity, rounded edges, and no internal structure are considered cysts and classified as "potential transmission risk." S5-2, concentration warning mechanism: if the live insect concentration is greater than 1 / 10L, the system will automatically issue a red alarm and upload the information to the supervision system API; Conduct event trend analysis, use the relationship between live insect concentration and time, and use the exponential model to determine whether the pollution is sudden or continuous.

7. The method for detecting and identifying two insects in water based on deep learning according to claim 1, characterized in that: In step S6, the following sub-steps are also included: S6-1 deploys the TensorRT inference engine, supports FP16 quantization and Layer Fusion, and increases inference speed to over 30FPS; it also supports dynamic input switching to adapt to different computing power; S6-2 crops the detection area and compresses it using JPEG 2000, transmitting only the valid area, reducing bandwidth usage by more than 50%. It also supports LoRaWAN remote upload, uploading insect count, activity ratio, and warning status in a package every hour. S6-3, the test results are uploaded to the Ethereum private chain through SHA-256 hash encryption to ensure data security and traceability, and the number of times the sample passes is recorded. If it exceeds 50 times, it will automatically prompt cleaning to avoid cross contamination caused by residues.

Images