Chironomidae larva recognition system based on deep learning and image recognition

Abstract

The invention relates to the field of biological recognition and environmental monitoring, and discloses a chironomidae larva recognition system based on deep learning and image recognition, which comprises an image acquisition module, an automatic sample processing device, an image superposition processing module, a deep learning recognition module and an automatic category classification and statistics module, the image acquisition module acquires multi-focal plane chironomidae larva images through an optical microscope and a digital camera, the automatic sample processing device realizes automatic loading operation of a sample, and the image superposition processing module performs adaptive fusion on the multi-focal plane images to generate a full-focal plane image containing multiple pieces of depth information; the deep learning recognition module is used for extracting fine structure features of chironomidae larvae and automatically recognizing and classifying different types of chironomidae larvae, and the automatic type classification and statistics module is used for performing statistical analysis according to classification results and generating a report of type distribution and time sequence change. The method is suitable for automatic monitoring of water ecological environment.

Description

Technical Field

[0001] The present invention relates to the technical field of biometric identification and environmental monitoring, and in particular to a chironomid larvae identification system based on deep learning and image recognition. Background Art

[0002] In the field of biological monitoring and environmental monitoring, species identification of chironomid larvae is one of the important indicators for assessing the health status of aquatic ecosystems. At present, traditional identification of chironomid larvae mainly relies on manual operations, including manual processing of samples, microscopic observation, and classification based on morphological characteristics. This method is not only time-consuming and labor-intensive, but also requires high professional knowledge and experience of operators, and is prone to subjective errors, making it difficult to ensure identification accuracy and efficiency. In addition, due to the tiny size of chironomid larvae, their fine structures (such as chin plates, tail setae, etc.) are difficult to fully present in a single focal plane. Traditional methods find it difficult to capture all key features, resulting in inaccurate identification.

[0003] Existing automated image recognition technology has been applied in some fields, but in the identification of chironomid larvae, due to the complex sample structure and unstable image quality, the existing technology has not been able to effectively combine multi-focal plane image processing and deep learning algorithms, and it is difficult to simultaneously solve the problems of sample processing automation, image clarity and recognition accuracy. Therefore, the existing technology still has many deficiencies in processing efficiency, recognition accuracy and degree of automation, and it is difficult to meet the needs of rapid and accurate recognition in water ecological monitoring.

[0004] To address these problems, the present invention proposes a Chironomid larvae identification system based on deep learning and image recognition technology, aiming to overcome the limitations of traditional methods, realize the automation and intelligence of sample processing, image acquisition and identification processes, and improve the accuracy and efficiency of identification. Summary of the invention

[0005] In response to the shortcomings of the prior art, the present invention provides a chironomid larvae identification system based on deep learning and image recognition, which solves the problems of low sample processing efficiency, poor recognition accuracy and difficulty in automation in traditional chironomid larvae identification methods, and realizes the intelligence and efficiency of sample processing, image acquisition and classification and recognition.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a chironomid larvae identification system based on deep learning and image recognition, comprising:

[0007] An image acquisition module includes an optical microscope with a transmitted light source, a microscope objective lens and a digital camera, wherein the digital camera is installed on the microscope and connected to a computer via a data transmission interface; the image acquisition module is controlled by computer software and automatically acquires multi-level images of chironomid larvae at different focal planes;

[0008] An automated sample handling device, comprising a sample loading system and a three-axis motorized stage, wherein the sample loading system transports the chironomid larvae sample to an imaging position, and the three-axis motorized stage adjusts the XYZ axis position of the sample for automatic image acquisition in multiple fields of view and multiple focal planes;

[0009] An image stacking processing module, wherein the image stacking processing module uses adaptive multi-scale image fusion to perform high-precision image stacking on the acquired images of different focal length planes to generate a full-focus plane image containing multiple depth information from bottom to top;

[0010] The deep learning recognition module, based on the convolutional neural network model, automatically identifies the microstructural features of the chironomid larvae, including but not limited to the chin plate, inner labial comb, eyespots and caudal setae, by training and extracting features from the full-focus images generated after superposition, so as to achieve the classification of chironomid larvae of different subfamilies and genera;

[0011] The module for automatic species classification and statistics automatically identifies and classifies the species of chironomid larvae by combining the output of the deep learning model, and provides statistical analysis. The results are stored in a database to support the needs of automated monitoring of biological monitoring and aquatic ecological environment.

[0012] The present invention also provides a method for identifying chironomid larvae based on deep learning and image recognition, comprising the following steps:

[0013] An image acquisition step, obtaining multi-focal plane images of the chironomid larvae through an image acquisition module, wherein the image acquisition module includes a transmitted light source, an optical microscope and a high-resolution digital camera, and controls the digital camera to automatically acquire image sequences at different focal planes;

[0014] The image stacking processing step is to perform image stacking processing on the acquired multi-focal plane images, and the image stacking processing includes:

[0015] Each focal plane image is decomposed into multiple scale levels through multi-scale Laplacian pyramid decomposition;

[0016] The fusion weight is calculated according to the local gradient, contrast and motion compensation results of the image, and multiple focal plane images are fused to generate a full focal plane image;

[0017] A feature extraction step, inputting the superimposed full-focus surface image into a deep learning recognition module, wherein the deep learning recognition module extracts features of the chironomid larvae based on a convolutional neural network, and the feature extraction process includes convolution, activation, and pooling operations;

[0018] In the classification step, the extracted features are classified through the fully connected layer and Softmax classifier of the deep learning recognition module, and the species label of the midge larvae and the corresponding classification confidence are output;

[0019] a classification result storage step, storing the classification result together with the image number, timestamp, and detection position information into a database;

[0020] In the statistical analysis step, the classification results are statistically analyzed through the automatic species classification and statistical module, the number of different species of chironomid larvae and their time series changes are calculated, the species distribution and time trend graphs are generated, and a report is automatically generated.

[0021] The present invention provides a chironomid larvae identification system based on deep learning and image recognition.

[0022] Beneficial effects:

[0023] 1. The present invention generates a full-focal-plane image through multi-focal-plane image acquisition and image superposition processing technology, which effectively solves the problem that a single focal-plane image cannot capture all the key structures of the chironomid larvae, and significantly improves the recognition accuracy of the microscopic structure of the chironomid larvae.

[0024] 2. The present invention adopts an automated sample processing device to realize automatic loading, positioning, pressing and image acquisition of samples, reducing the errors caused by manual operation, ensuring the stability of samples and the consistency of image acquisition, and improving the stability of the recognition system.

[0025] 3. The present invention uses a deep learning recognition module and a convolutional neural network to automatically extract features from the full-focus image and efficiently classify the chironomid larvae. Compared with the traditional manual classification method, deep learning can classify species more quickly and accurately, greatly improving processing efficiency.

[0026] 4. The present invention screens the classification results by confidence, ensuring that only high-confidence classification results are used for automatic classification and statistics, thereby improving the reliability of the system classification results. At the same time, for low-confidence results, a manual review mechanism is provided to further improve classification accuracy.

[0027] 5. The automatic species classification and statistics module of the present invention can perform species quantity statistics and time series analysis in real time, and display the dynamic changes of the chironomid larvae population. This provides timely and accurate data support for long-term monitoring and research of the water ecological environment, and facilitates environmental change analysis and trend prediction.

[0028] 6. The system of the present invention can automatically generate reports containing classification results, statistical analysis, and time series change trends, and supports export in multiple formats, which is convenient for users to archive and analyze, reduces the workload of manual data sorting and report writing, and improves the degree of automation of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1It is a schematic diagram of the system architecture of the present invention;

[0030] Figure 2 It is a schematic diagram of the structure of the image acquisition module and the automatic sample processing device of the present invention;

[0031] Figure 3 It is a schematic diagram of the structure of the glass slide tank of the present invention;

[0032] Figure 4 It is a computer control connection diagram of the present invention;

[0033] Figure 5 It is a schematic diagram of a picture superposition processing module of the present invention;

[0034] Figure 6 A schematic diagram of a deep learning recognition module of the present invention;

[0035] Figure 7 This is a schematic diagram of the automatic classification and statistics module of the present invention;

[0036] Figure 8 It is a schematic diagram of the method flow of the present invention.

[0037] Among them, 10, optical microscope; 11, transmitted light source; 12, objective lens; 13, stage; 14, digital camera; 15, three-axis electric stage; 20, computer; 30, automated sample processing device; 31, sample bottle; 32, sample injection hose; 33, peristaltic pump; 34, sample delivery hose; 35, sample tank; 36, sample discharge hose; 37, waste liquid bucket; 40, slide tank; 50, image overlay processing module; 51, multi-focal plane image acquisition unit; 52, multi-scale image decomposition unit; 53. Adaptive weighted fusion unit; 54. Image fusion unit; 55. Image reconstruction unit; 60. Deep learning recognition module; 61. Input preprocessing unit; 62. Convolution feature extraction unit; 63. Feature classification unit; 64. Training and optimization unit; 65. Database interface unit; 70. Category automatic classification and statistics module; 71. Classification result storage unit; 72. Confidence screening unit; 73. Category statistics unit; 74. Time series analysis unit; 75. Report generation unit. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 creative work are within the scope of protection of the present invention.

[0039] Please see attached Figure 1 -Attached Figure 7 The present invention provides a chironomid larvae identification system based on deep learning and image recognition, which can realize the automatic identification, classification and statistical analysis of chironomid larvae. The system realizes efficient and accurate detection and identification of chironomid larvae by combining automated sample processing, image acquisition, image overlay processing, deep learning recognition, and automatic classification of species with statistical analysis modules, providing technical support for biological monitoring and water ecological environment monitoring. The specific embodiments of the present invention are described in detail below:

[0040] The image acquisition module is one of the core components of the entire system, used to obtain multi-focal plane images of chironomid larvae. The module combines sophisticated optical imaging equipment with a high-resolution digital camera to ensure high-precision imaging of the microstructure of chironomid larvae, providing high-quality image data for subsequent image processing and deep learning recognition.

[0041] The optical microscope 10 is the core imaging device in the image acquisition module, which is used to observe the chironomid larvae samples at high magnification. The optical microscope is mainly composed of a transmitted light source 11, an objective lens 12 and a stage 13, which ensures the visibility and high-precision imaging of the sample. The specific description is as follows:

[0042] Transmitted light source 11: used to provide uniform and stable illumination, ensuring the formation of transmitted light under the sample, so that each microstructure of the sample is clearly visible. The transmitted light source 11 can adjust the brightness according to the needs of different samples to avoid affecting the image quality due to too strong or too weak light.

[0043] Objective lens 12: The optical microscope 10 is equipped with objective lenses 12 of various magnifications, such as 10x, 40x, and 100x objective lenses, and the user can select a suitable objective lens magnification according to the sample size. The objective lens 12 has high optical transmittance and can magnify the tiny structures of the chironomid larvae, such as the chin plate, caudal setae, and inner labial comb, to ensure clear features.

[0044] The digital camera 14 is installed above the objective lens of the microscope to capture the sample image imaged by the optical microscope. The camera has a high resolution, usually 20 million pixels or more, and can record detailed images of chironomid larvae in high definition. Its specific functions are as follows:

[0045] Resolution: High-resolution cameras ensure that every subtle feature of the chironomid larvae is captured, especially at high magnification, ensuring the image is clear and undistorted.

[0046] Autofocus function: The camera is equipped with an autofocus function, which can automatically adjust the focus when shooting in multiple focal planes to ensure that the sample is in the best imaging state in each focal plane to avoid blurred images.

[0047] High frame rate shooting: The camera supports high frame rate continuous shooting mode, which can complete continuous image acquisition of different focal planes in a short time and improve work efficiency. It can shoot more than 5 frames per second to ensure the integrity of the sample.

[0048] Data transmission interface: The digital camera 14 is connected to the computer via a high-speed data transmission interface (such as USB 3.0 or higher standard interface) to transmit the collected image data in real time to ensure efficient data processing. The data transmission speed can reach more than 5Gbps to ensure the stability of image transmission.

[0049] The computer 20 is the core control unit of the image acquisition module and is responsible for managing the entire image acquisition process. The computer has built-in dedicated control software that can interact with the digital camera 14 and the optical microscope 10, automatically control various parameters in the image acquisition process, and store and manage the acquired images. Its functions include:

[0050] Automatic control: By presetting acquisition parameters, the computer 20 can automatically control the camera's exposure, focus, shooting speed and other operations, reducing manual intervention and improving accuracy and efficiency.

[0051] Real-time image display: The computer can display images in real time during the acquisition process, allowing the operator to monitor imaging quality and make necessary parameter adjustments.

[0052] Image storage: The acquired images are automatically saved to local or network storage devices, and the file name is named with a timestamp or sample number for easy subsequent retrieval and processing. The system supports multiple image formats, including RAW, JPEG and TIFF.

[0053] The multi-focal plane automatic control unit is connected to the computer 20 and the three-axis electric stage 15, and is responsible for automatically collecting images in multiple focal planes. By adjusting the Z axis of the microscope, images of the sample at different focal lengths can be continuously collected. The functions of the multi-focal plane automatic control unit are as follows:

[0054] Automatic Z-axis adjustment: This unit can accurately adjust the position of the sample on the Z-axis, so that the focus of the image collected at different focal lengths is different, thereby obtaining information at different depths of the sample. The adjustment range can be precisely controlled at the micron level, with each step moving 1-2 microns to ensure that multi-level detailed images are collected.

[0055] Focus range setting: Users can pre-set the focus range and shooting interval through the software interface. The system automatically adjusts the focus according to these settings and completes the image acquisition. Typically, the depth range of 0 to hundreds of microns can be covered to ensure that all levels of information of the sample are fully recorded.

[0056] The XYZ three-axis motorized stage 15 is used to fix the sample and make precise adjustments in the XYZ directions. In image acquisition, the stage mainly performs the following functions:

[0057] X and Y movement: The stage can be moved in the horizontal direction (X and Y) to ensure that the entire sample area is scanned. For large samples, multi-field image stitching technology can be used to complete the scanning of the entire area.

[0058] Z-axis movement: In conjunction with the multi-focal plane automatic control unit, the stage accurately adjusts the sample height along the Z-axis direction, allowing the camera to capture images at different focal planes. The Z-axis adjustment accuracy reaches the micron level, ensuring that each focal plane can clearly present the structural information of the sample.

[0059] The software control interface is the user interaction platform of the computer system and is responsible for the visual control of the entire image acquisition process. The operator can complete a series of acquisition settings and operations through the interface, including:

[0060] Exposure and focus settings: The operator can manually or automatically adjust the camera's exposure time, focus, light brightness and other parameters to ensure the best quality of image acquisition.

[0061] Acquisition mode selection: The system supports multiple image acquisition modes, including single-view multi-focal plane acquisition and multi-view scanning acquisition. Users can choose the appropriate mode according to their needs.

[0062] Image preview and monitoring: Preview the captured images in real time and adjust the capture parameters in time to avoid capture errors due to image blur or improper exposure.

[0063] The overall process of the image acquisition module:

[0064] 1. The operator sets the acquisition parameters through the software control interface, including focal range, exposure time and shooting interval.

[0065] 2. The sample is adjusted to an appropriate position by the XYZ three-axis electric stage 15, and the optical microscope 10 cooperates with the transmitted light source 11 to provide stable illumination.

[0066] 3. The digital camera 14 starts to collect multi-focal plane images at different focal lengths, and the images are transmitted to the computer 20 in real time through the data transmission interface.

[0067] 4. The computer 20 manages and stores the acquired images, and the operator can monitor the acquisition process in real time through the software interface to ensure image quality.

[0068] The image acquisition module realizes multi-level and high-precision image acquisition of chironomid larvae through the organic combination of a precise optical microscope 10, a high-resolution digital camera 14, a three-axis electric stage 15 and a computer 20. Each component cooperates with each other to ensure that the complete depth information of the sample is obtained at different focal planes, providing high-quality image data for subsequent image processing and recognition.

[0069] The automated sample handling device 30 is an important component of the system, responsible for the transmission, positioning and fixation of the chironomid larvae samples to ensure that the samples are in a stable state during the microscope imaging process and can be accurately collected in multiple fields of view and multiple focal planes. Since the body length of each chironomid larva is usually 1-2 cm, the system will perform a certain number control in a single field of view to ensure the stable positioning and imaging quality of the sample. Through the automated adjustment of the sample position and a variety of processing operations, errors in manual operations are avoided, greatly improving the accuracy and efficiency of the imaging and recognition process.

[0070] The sample loading system is responsible for automatically transporting the chironomid larvae samples from the storage container to the imaging area, and ensuring that the number of samples in the field of view is controlled to 1 each time the image is taken, so as to avoid multiple samples interfering with the imaging quality. The specific structure and functions include:

[0071] Sample injection device: The chironomid larvae sample is transported from the sample bottle to the imaging position through the injection hose 32. The system has a sample flow control function, which can accurately adjust the sample injection speed to ensure that only one chironomid larvae individual is included in the field of view each time. The injection hose 32 is connected to the sample liquid bottle 31 and the peristaltic pump 33, and the sample is transported to the imaging position above the stage under controlled liquid flow.

[0072] Peristaltic pump 33: The peristaltic pump 33 is used to accurately control the flow rate of the liquid sample to prevent the sample from flowing too fast and affecting the positioning or imaging of the sample. Through the adjustable control of the peristaltic pump 33, the system can adjust the liquid flow rate according to the type and density of the sample to ensure that the sample can be stably and continuously transferred to the imaging area. The peristaltic pump head applies appropriate squeezing to the inlet hose 32 to promote the flow of liquid without direct contact with the liquid, thereby avoiding contamination or sample damage. In addition, since the body length of the chironomid larvae is 1-2 cm, the peristaltic pump can control the sample injection speed according to preset parameters to ensure that only one complete individual is captured at a time within the field of view. The peristaltic pump controls the sample flow at a lower speed through the hose to prevent multiple samples from entering the imaging area.

[0073] The sample processing groove 35 is the key area for sample flow and collection on the stage. The groove is connected to the sample delivery hose 34 through a peristaltic pump 33 to inject the chironomid larvae sample, which is convenient for microscope photography and detection. The groove is injected with chironomid larvae samples through a peristaltic pump, and the flow rate and number of samples are controlled by software. The specific features of the sample processing groove include:

[0074] Liquid control: After the liquid sample is injected into the sample processing groove through the peristaltic pump 33, the system will adjust the liquid flow rate and liquid volume as needed to ensure that the sample maintains an appropriate observation state in the groove. Liquid control has a great influence on the stability of the sample, and the appropriate flow rate helps to correctly position the sample in the microscope field of view.

[0075] The waste liquid treatment system is used to collect waste liquid generated during the imaging process to ensure the cleanliness of the experimental environment and the safe handling of samples. Its functions include:

[0076] Waste liquid diversion device: The sample waste liquid is introduced into the waste liquid treatment system by connecting the sample discharge hose 36. The sample discharge hose 36 is connected to the sample tank 35 to guide the excess liquid into the waste liquid bucket 37 to avoid sample overflow or contamination of the stage.

[0077] Workflow of automated sample processing device:

[0078] 1. The sample loading system transfers the chironomid larvae sample from the sample storage bottle to the sample processing tank 35 through the peristaltic pump 33, ensuring that only one individual enters the imaging area at a time within the field of view.

[0079] 2. The sample is precisely moved in the XYZ direction by the three-axis electric stage 15, and automatically adjusted to different fields of view and focal planes, completing the image acquisition of the sample in multiple fields of view and multiple focal planes.

[0080] 3. The waste liquid generated during the collection process is discharged through the sample waste liquid treatment system to ensure the cleanliness of the experimental environment.

[0081] The automated sample processing device 30 realizes fully automated operation from sample loading to imaging by integrating a sample loading system, a peristaltic pump 33, a three-axis electric stage 15, and a waste liquid treatment system. It also ensures that only one individual is included in the field of view during each imaging process, avoiding interference of multiple samples with the image quality. The device ensures the stability and accuracy of the sample throughout the imaging process, avoids errors in manual operation, and provides a reliable sample processing basis for high-precision identification and classification of chironomid larvae.

[0082] In some embodiments, the upper surface structure of the three-axis electric stage 15 of the automated sample processing device 30 is optimized to meet the needs of different types of sample processing. In order to meet the special application scenarios of manual operation, especially when manual operation of chironomid larvae is required, the system adds a slide slot 40 for placing a slide, and provides a process design for manually sucking sample liquid, placing a slide, and pressing a slide. The following is a detailed description of this embodiment.

[0083] Structural design of slide trough

[0084] The slide groove 40 is located on the upper surface of the three-axis electric stage 15 and is used to stably place the slide required for the microscope. The slide groove is designed as a groove structure that conforms to the size of a standard microscope slide and has the following functions:

[0085] Stably place the slide: The slide groove 40 can fix the position of the slide to prevent the slide from sliding or shifting during the movement of the stage, ensuring that the sample remains stable during the pressing and imaging process.

[0086] In some operations, the user can manually aspirate the sample solution in the sample processing tank 35 and drop it onto the surface of the slide. This operation is suitable for sample solution processing scenarios that require manual control, and the steps are as follows:

[0087] 1. Sample liquid absorption: The user absorbs an appropriate amount of sample liquid from the sample processing tank 35 through a manual pipette. The capacity of the manual pipette is reasonably designed to absorb an appropriate amount of liquid at one time to avoid excessive sample liquid dripping on the glass slide.

[0088] 2. Slide placement: The user places a microscope slide of standard size into the slide groove 40. The slide groove 40 has an appropriate depth and border to ensure that the slide is firmly placed in the groove.

[0089] 3. Sample liquid dripping: The user drips the absorbed sample liquid evenly on the glass slide. The sample liquid is evenly distributed on the surface of the glass slide, which is conducive to the fixation of the sample and the subsequent pressing operation.

[0090] Manual tableting process of sample solution

[0091] After the sample liquid is dripped onto the glass slide, the system will use the tablet pressing device to press the sample liquid onto the slide, so that the chironomid larvae sample is evenly and stably attached to the glass slide. The specific process of manual tablet pressing is as follows:

[0092] Sample spreading: After the sample solution is dropped onto the slide, the chironomid larvae sample is distributed on the slide surface. Users can swing or tilt the slide appropriately to make the sample more evenly distributed in the center area of ​​the slide, which is convenient for subsequent observation and imaging.

[0093] Manual pressing: Users can adjust the pressure of the pressing device through the system or manually to press the sample. The pressing device applies uniform pressure to make the chironomid larvae sample adhere to the glass slide evenly. Especially for tiny structures such as the chin plate and tail setae, the pressing operation helps to show these key parts of the sample more clearly.

[0094] Identification process after sample hydraulic sheet

[0095] After manual pressing is completed, the system will automatically enter the image acquisition and recognition stage. At this point, the sample on the slide has been fixed and flattened, ensuring high-resolution microscopic imaging and identification of the sample's microstructure. The specific identification process is as follows:

[0096] Automatic positioning: The slide placed in the slide slot will enter the field of view of the microscope as the three-axis motorized stage 15 is adjusted in the XYZ direction. The system ensures that the key area of ​​the sample is in the center of the microscope lens through precise movement of the XYZ axis.

[0097] Multi-focal plane image acquisition: The system acquires multi-focal plane images of the pressed sample by combining a digital camera 14 and an optical microscope 10. By adjusting the Z-axis direction, the system can obtain image information of the sample at different depths, ensuring that all key structures of the sample (such as the chin plate, caudal setae, etc.) can be clearly imaged within the focus range.

[0098] Deep learning recognition: The collected images are processed through the image processing module and the deep learning recognition module for feature extraction and automatic classification. The deep learning model can automatically classify the species and subfamily of chironomid larvae according to the different microscopic structures of the samples, and store the results in the database.

[0099] Advantages and application scenarios of manual operation

[0100] The manual aspiration and tableting operation provides flexibility for certain special experimental requirements. The following are the advantages and applicable scenarios of this implementation:

[0101] Flexible operation: Through manual aspiration and tableting, users can finely control sample processing, which is especially suitable for the processing of special or precious samples, avoiding sample damage or errors that may occur during the fully automated process.

[0102] Wide adaptability: This manual operation process is suitable for scenarios where special treatment of samples is required, such as when different fixatives or stains need to be added to different samples. Manual operation makes sample processing more flexible.

[0103] Experimental optimization: In scenarios where samples need to be repeatedly processed, sample concentrations adjusted, or different experimental conditions tested, manual operations can be used in conjunction with automated devices to optimize experimental results.

[0104] This embodiment provides a flexible manual operation process for the system, and specially designs the functions of the slide slot, manual pipette and slide pressing operation, so that the user can manually draw the sample liquid, drip the slide and perform the slide pressing operation. This operation mode provides more controllability and flexibility for the sample processing process, and is suitable for experimental scenarios that require fine control.

[0105] The image stacking processing module 50 is one of the key components of the system, responsible for fusing and processing the data from the multi-focal plane images to generate a high-resolution image with a full focal depth. Through the processing of this module, the images of the chironomid larvae samples at different depths can be accurately aligned and stacked to generate an image with a full focal plane. This not only greatly improves the image quality, but also ensures that the key features of the chironomid larvae, such as the chin plate, inner labial comb, tail bristles, etc., can be clearly displayed in a single image.

[0106] The multi-focal plane image acquisition unit 51 acquires images of the chironomid larvae at different focal length planes by combining the XYZ three-axis electric stage 15 and the digital camera 14. Each set of image sequences contains information at different depths, and can capture the microstructure of the chironomid larvae at different focal lengths.

[0107] Multi-focal plane imaging: The digital camera 14 works with the microscope. By adjusting the height of the Z axis, the system can continuously shoot at different focal planes. Each image corresponds to the layer of the sample at different depths. The image acquisition process has a high frame rate, ensuring that a large number of image sequences are collected in a short time, avoiding image blur caused by sample movement or liquid interference.

[0108] Data synchronization and caching: During the acquisition process, the system caches the acquired images into the memory in real time and transmits them to the processing module through the high-speed data interface. Image data transmission and processing are carried out in parallel to ensure the efficiency and real-time performance of the system.

[0109] The multi-scale image decomposition unit 52 is used to decompose the multi-focal plane image, and decompose each image into image levels of different scales by using the Laplacian pyramid or Gaussian pyramid method. The purpose of this step is to extract different frequency information of the image and provide data support for subsequent image fusion.

[0110] Laplacian pyramid decomposition: This unit uses the Laplacian pyramid algorithm to decompose each focal plane image into multiple scale levels. First, the image is passed through a Gaussian filter to generate a Gaussian pyramid, and then the difference between adjacent scale images is calculated to generate a Laplacian pyramid. The formula is as follows:

[0111]

[0112] in, represents the Laplacian pyramid representation of the i-th focal plane image at the l-th scale, is the Gaussian pyramid representation of the image.

[0113] Separation of high-frequency and low-frequency features: Through image decomposition, the low-frequency (smooth areas) and high-frequency (edge ​​and texture details) information of the image can be effectively separated. The high-frequency part usually represents the details and sharp edges in the image, while the low-frequency part represents the illumination and a wide range of tonal changes. This separation provides the basis for subsequent image fusion, so that images in different focal planes can be optimized and fused according to the clarity of specific areas.

[0114] The adaptive weighted fusion unit 53 dynamically adjusts the fusion weight of each focal plane image according to the local clarity, contrast and motion compensation result of each focal plane image to ensure that the key detail area is best displayed in the final full-focus plane image.

[0115] Local gradient calculation: This unit evaluates the clarity of the image by calculating the local gradient of each focal plane image. The larger the gradient, the richer the edge and texture information of the image, indicating that the area is clearer in the focal plane image. The gradient calculation formula is as follows:

[0116]

[0117] in, is the gradient value of the i-th focal plane image at pixel x, y.

[0118] Local contrast calculation: In addition to gradient, the system also evaluates the clarity of the image by calculating the local contrast of the image. The local contrast reflects the brightness difference in a certain area of ​​the image and can reveal whether the image has enough visual information.

[0119] Motion compensation: If the sample or environment moves slightly (such as liquid disturbance) during multi-focal plane image acquisition, the system corrects these movements through motion compensation technology. Through optical flow or phase correlation, the system can estimate the motion vector of the image and compensate for it. The formula is:

[0120]

[0121] Among them, v(x,y) is the motion vector of the sample at time t, I(x,y,t) is the image grayscale value at time t, and the motion vector is used for correction before image superposition.

[0122] Weighted fusion formula: The adaptive weighted fusion unit calculates the weighted value of each focal plane image based on the local gradient, contrast and motion compensation results. The specific calculation formula is:

[0123]

[0124] Among them, w i (x,y) is the weight of the i-th focal plane image, is the local gradient of the image, C i (x, y) is the local contrast, M i (x, y) is the image quality weight after motion compensation, and α and β are adjustment coefficients.

[0125] The image fusion unit 54 is responsible for fusing the images of different scales after weighted calculation. For each decomposition scale, the system fuses the Laplacian pyramid layers of each focal plane image according to the weighted value to generate a full focal plane image containing multiple layers of information.

[0126] Multi-scale fusion process: At each scale, the final fused image is calculated based on the weighted values ​​of each focal plane image. The fusion formula is as follows:

[0127]

[0128] in, is the fused image at the lth scale.

[0129] Layer-by-layer fusion: The fusion process starts from the lowest scale and gradually expands to high-frequency information, ensuring that the clearest image content is retained in each frequency band. In the low-frequency scale, more light and color information is fused, while in the high-frequency scale, the system pays more attention to details such as edges and textures.

[0130] The image reconstruction unit 55 is responsible for recombining the fused multi-scale images into a complete full-focus image. Through the operation of the inverse Laplacian pyramid, the system can restore the size and resolution of the original image and generate a high-resolution image with a complete focal depth.

[0131] Image reconstruction process: The final full-focus image is reconstructed by inverse transforming the fused Laplacian pyramid layer by layer. The reconstruction process is:

[0132]

[0133] Among them, I fused (x, y) is the final full-focal plane image after fusion, and L is the number of decomposed scales.

[0134] Feature enhancement: During the image reconstruction process, the system can also enhance the features of key structures (such as the chin plate, inner labial comb and tail bristles of the midge larvae). By locally enhancing high-frequency information, the system ensures that these tiny structures are clearly displayed in the final image.

[0135] Prioritized fusion of feature regions

[0136] The system optimizes the specific structural areas of the chironomid larvae. For key feature areas such as the chin plate and tail bristles, the system assigns higher weights to these areas during the fusion process, making them more prominent in the full-focus image, thereby improving the recognition accuracy of the subsequent deep learning model.

[0137] The workflow of the image overlay processing module 50 is as follows:

[0138] 1. Multi-focal plane image acquisition: The system uses a three-axis motorized stage and a digital camera to collect images at multiple depths at different focal lengths to form an image sequence.

[0139] 2. Multi-scale image decomposition: Perform multi-scale decomposition on each image to generate Laplacian pyramid and Gaussian pyramid to separate image information of different frequencies.

[0140] 3. Weighted fusion: Dynamically adjust the fusion weight of each focal plane image through local gradient, contrast and motion compensation results, and perform multi-scale fusion.

[0141] 4. Image reconstruction: Through the inverse Laplace pyramid process, the multi-scale image is reconstructed into a complete full-focus image to ensure that all key structures are clearly displayed.

[0142] The image overlay processing module 50 generates a high-quality full-focal plane image through multi-focal plane image acquisition, decomposition, weighted fusion and reconstruction, ensuring that the structural features of the chironomid larvae sample at different depths can be clearly presented at the same time. The multi-scale fusion and feature area priority processing strategy of this module not only improves the overall quality of the image, but also provides high-quality data input for subsequent deep learning recognition, improving the recognition accuracy and efficiency of the system.

[0143] The deep learning recognition module 60 is the intelligent core of the system, responsible for automatically recognizing and classifying the full-focal-surface images generated by the image stacking processing module. Through the multi-level feature extraction capability of the convolutional neural network (CNN), the module can recognize the fine structure of the chironomid larvae and achieve accurate species classification based on their characteristics. In order to improve the robustness and accuracy of the system, the deep learning recognition module combines multiple submodules such as data preprocessing, feature extraction, classification prediction, model training, and optimization algorithms.

[0144] The input preprocessing unit 61 is responsible for data cleaning and preprocessing of the full-focus image output by the image stacking processing module. In order to meet the input requirements of the deep learning model, the image needs to be normalized, enhanced, and resized.

[0145] Image normalization: Normalizing image pixel values ​​to the [0, 1] range helps speed up the model training process and improve the convergence of the model.

[0146] Image resizing: To adapt to the input layer size of the convolutional neural network, the image needs to be resized. Common input sizes are 224×224 or 512×512. The system uses bilinear interpolation to scale the image to ensure that detail information is retained during the scaling process.

[0147] Data enhancement: During the model training phase, data enhancement techniques can generate more diverse samples and enhance the robustness of the model. Common data enhancement operations include:

[0148] Random rotation: Randomly rotate the image in the range [-15°, +15°] to prevent the model from being overly dependent on a specific orientation.

[0149] Random Crop: Randomly crop a portion of the image to simulate the diversity of image data.

[0150] Random flip: Flip the image horizontally or vertically to enhance the model’s generalization ability for samples at different positions.

[0151] The convolution feature extraction unit 62 is the core part of the deep learning recognition module, responsible for extracting high-level feature representations from the input full-focus image. This unit is based on the convolutional neural network (CNN) architecture, and gradually extracts the microscopic structural features of the midge larvae, such as the chin plate, caudal setae, and inner labial comb, through multiple convolutional layers and pooling layers.

[0152] Convolutional layer: The convolutional layer performs sliding convolution operations on the image through multiple convolution kernels (usually 3×3 or 5×5 convolution kernels) to extract local features of the image and generate feature maps. The formula for the convolution operation is as follows:

[0153]

[0154] Among them, I normalized (x, y) is the normalized full-focal-plane image, w(i, j) is the weight of the convolution kernel, F(x, y) is the convolution result, and b is the bias.

[0155] Activation function: After each convolution layer, a nonlinear activation function (such as ReLU) is used to process the convolution result to enhance the model's ability to express nonlinear features. The activation formula of ReLU is:

[0156] A(x,y)=max(0,F(x,y))

[0157] Among them, A(x,y) is the output after activation.

[0158] Pooling layer: The pooling layer is used to reduce the dimension and downsample the activated feature map to reduce the amount of calculation and enhance the spatial invariance of the feature. The commonly used pooling operation is maximum pooling, and the formula is as follows:

[0159]

[0160] Where P is the size of the pooling window, usually 2×2.

[0161] Stacking of convolutional layers: Deep convolutional neural networks are usually composed of multiple stacked convolutional and pooling layers. As the number of layers increases, the network can extract more complex and higher-level features. For example, shallow convolutional layers may extract edge and texture information of chironomid larvae, while deep convolutional layers can capture more semantic features, such as the microscopic shapes of the chin plate and tail bristles.

[0162] The feature classification unit 63 is responsible for further processing the high-level features processed by the convolution feature extraction unit and outputting the classification results of the species of the midge larvae. The feature classification unit includes a fully connected layer and a Softmax classifier, and the specific steps are as follows:

[0163] Fully connected layer: The feature map output by the convolutional feature extraction unit will be flattened into a one-dimensional vector and passed to the fully connected layer to map the high-dimensional features to the classification space. The calculation formula of the fully connected layer is:

[0164] z=W·x+b

[0165] Among them, W is the weight matrix of the fully connected layer, x is the input feature vector, b is the bias, and z is the feature representation of the output.

[0166] Softmax classifier: The Softmax classifier is used to convert the output of the fully connected layer into the probability distribution of each type of chironomid larvae. The formula is as follows:

[0167]

[0168] Among them, z j is the score of category j, N is the number of chironomid larvae categories, and p(y=j|x) is the probability that the input image belongs to category j.

[0169] The training and optimization unit 64 is used for model training and parameter optimization. The weights in the network are updated through the back propagation algorithm so that the model can better learn the characteristics of the chironomid larvae. This unit includes the following key parts:

[0170] Loss function: The loss function of the model uses cross entropy loss to measure the difference between the model prediction result and the true label. The cross entropy loss is defined as:

[0171]

[0172] Among them, y i is the true label of the sample, p i is the predicted probability output by the model.

[0173] Optimizer: To accelerate model convergence, the system uses the Adam optimizer, which is an optimization algorithm with adaptive learning rate.

[0174] Back propagation: In each iteration, the system calculates the gradient through the back propagation algorithm and uses the optimizer to update the network weights to reduce the value of the loss function. Back propagation ensures that the network can continuously correct errors, making the model's prediction results more accurate.

[0175] The database interface unit 65 is responsible for storing the classification results and related information (such as image number, timestamp, classification confidence, etc.) in the database to facilitate subsequent statistical analysis and data query. This unit has the following functions:

[0176] Classification result storage: The result of each classification, including the type of chironomid larvae, confidence, and related metadata (such as detection time and sample location), will be stored in the database.

[0177] Query and update: The system supports query and update of historical classification results. Users can retrieve classification records of certain samples or update incorrect classification results as needed.

[0178] The workflow of the deep learning recognition module:

[0179] Input preprocessing: The full-focus image is normalized and data enhanced through the input preprocessing unit 61 to ensure that the image is adapted to the input requirements of the convolutional neural network.

[0180] Feature extraction: The convolutional feature extraction unit 62 uses multiple convolutional layers and pooling layers to extract high-level features of the midge larvae layer by layer.

[0181] Feature classification: Use the fully connected layer and Softmax classifier in the feature classification unit 63 to map the extracted features to the classification space, and output the type of chironomid larvae and the classification confidence.

[0182] Training and optimization: The network is trained through the training and optimization unit 64, and the model weights are updated using backpropagation and the Adam optimizer to continuously improve the classification accuracy of the model.

[0183] Result storage: The classification results are stored in the database through the database interface unit 65 to facilitate subsequent query and statistical analysis.

[0184] The deep learning recognition module can accurately identify the microstructure characteristics of chironomid larvae and achieve efficient species classification through the organic combination of data preprocessing, convolution feature extraction, classification prediction and optimization training. This module has powerful feature extraction and classification capabilities, and improves the accuracy and robustness of classification by continuously optimizing model weights. The database interface unit ensures the storage and query of classification results, providing data support for subsequent statistical analysis.

[0185] The automatic species classification and statistics module 70 is the core functional module in the entire system that is responsible for receiving the output of the deep learning recognition module, classifying and statistically analyzing the chironomid larvae. This module combines the classification results of the deep learning model to automatically classify and count different types of chironomid larvae, provide real-time species analysis and dynamic change monitoring, and generate relevant reports. Through this module, users can fully understand the distribution and quantity change trends of chironomid larvae, and provide data support for water ecological environment monitoring and biodiversity research.

[0186] The classification result storage unit 71 is responsible for storing the classification results output by the deep learning recognition module into the database and associating them with metadata such as image number, timestamp, location, etc. This unit ensures the persistence and retrievability of the classification results, facilitating subsequent query and analysis.

[0187] Classification result record: The classification result includes the species label, confidence level and other information of the chironomid larvae. The classification result of each sample will be stored in the database together with the corresponding image ID, detection time, sampling location and other metadata to generate a complete record entry. The entry is stored in a structured manner to facilitate quick query by the system.

[0188] Metadata association: In addition to the classification label, the system automatically associates relevant information about the detection, such as image number, timestamp, detection location, etc., to the classification result. These metadata can provide contextual information about sample collection, especially in environmental monitoring tasks, and can be analyzed based on different sampling points and times.

[0189] Data retrieval and update: The system supports users to query and update classification results. If the recognition results need to be manually adjusted or confirmed, users can edit and update the classification results of specific samples through the query system to ensure the accuracy of the data.

[0190] The confidence screening unit 72 is used to screen the classification results according to the confidence of the deep learning model output to ensure the reliability of the classification results. For samples with low confidence, the system will mark them as uncertain classifications and allow users to manually review or further process them.

[0191] Confidence threshold setting: This unit allows users to set the classification confidence threshold. Classification results with low confidence will be marked as "uncertain" and can be processed through further manual review or model optimization. For example, when the confidence of the classification result is lower than the set threshold pthresholdpthreshold, the system marks the result as uncertain, and the screening process is defined as:

[0192]

[0193] Among them, y i is the classification label, p i is the classification confidence, p threshold The confidence threshold is preset.

[0194] Manual review mechanism: For classification results marked as uncertain, the system provides a manual review function, which allows users to manually review sample images and manually adjust classification results as needed. This mechanism helps improve the overall accuracy of classification results.

[0195] The species statistics unit 73 is used to perform statistical analysis on the stored classification results and provide the distribution of various types of chironomid larvae in the test samples. This unit can quickly calculate the number ratio of different species and generate corresponding statistical charts.

[0196] Quantity statistics: The system counts the number of different types of chironomid larvae based on the classification results, and calculates the number of each type of larvae in the test sample and its proportion. The statistical formula is as follows:

[0197]

[0198] Among them, 1(y j =y i ) is the indicator function, when y j =y i It returns 1 if the value is 0, otherwise it returns 0. m is the number of samples.

[0199] Statistical results display: The statistical unit can generate various types of charts, including pie charts, bar charts, etc., to show the distribution of the number of various types of chironomid larvae. These charts provide users with a clear visual data display, helping users to quickly understand the species composition of the current sample.

[0200] The time series analysis unit 74 is used to perform time series analysis on the classification results of the chironomid larvae, and to display the number change trend of various types of chironomid larvae in different time periods, which is particularly suitable for long-term monitoring tasks and environmental change analysis.

[0201] Time series data processing: The system groups and counts the number of each type of chironomid larvae by time according to the timestamp information in the classification results, and generates time series data for each type of chironomid larvae. Users can set time windows, such as day, week, month, etc., to flexibly analyze the dynamic changes of chironomid larvae populations in different time periods.

[0202] Trend analysis and visualization: The system displays the results of time series analysis in the form of line graphs, etc., reflecting the changing trend of the number of various types of chironomid larvae over time. Time series charts can clearly show the growth, decrease or stability of the population, helping users understand the impact of environmental changes on the chironomid larvae population.

[0203] The report generation unit 75 is responsible for automatically generating a detailed report file based on the classification results and statistical analysis. The report includes the species distribution of the chironomid larvae, time series analysis results and other statistical data, and supports exporting the report in multiple formats for users to archive or further analyze.

[0204] Report Content: The automatically generated report includes the following information:

[0205] Summary of classification results: Count the number and proportion of each type of chironomid larvae to provide the species composition of the current sample.

[0206] Temporal change trends: Through time series analysis results, the dynamic changes of various types of chironomid larvae in different time periods are displayed, reflecting the long-term change trend of the population.

[0207] Confidence Statistics: Displays the confidence distribution of various classification results to help users understand the reliability of classification results.

[0208] Report format export: The report generation unit supports exporting reports to multiple formats. Users can choose to save reports in PDF, Excel or CSV formats. PDF format reports are convenient for printing and archiving, while Excel and CSV formats are convenient for users to conduct further data analysis and processing.

[0209] The workflow of the automatic classification and statistics module 70 is as follows:

[0210] Classification result storage: The classification results of the deep learning recognition module 60 are stored in the database through the classification result storage unit 71 and associated with relevant metadata to ensure the integrity and traceability of the classification results.

[0211] Confidence screening and processing: The system screens the classification results through the confidence screening unit 72. For classification results with low confidence, the system marks them as "uncertain classification" and provides a manual review function.

[0212] Species statistics and display: The system counts the number of various types of chironomid larvae through the species statistics unit 73, and generates a species distribution chart for users to view intuitively.

[0213] Time series analysis: Through the time series analysis unit 74, the system displays the trend of the number of various types of chironomid larvae over time, helping users understand the dynamic changes of samples over time.

[0214] Report generation and export: The report generation unit 75 automatically generates a complete report file based on the classification results and statistical analysis, and supports export in multiple formats.

[0215] The automatic classification and statistics module provides powerful automated analysis capabilities and user interaction functions through automatic storage of classification results, confidence screening, species statistics and time series analysis. This module provides a complete solution for the automatic classification and statistical analysis of chironomid larvae species, which can meet the needs of multiple application scenarios such as environmental monitoring and biodiversity research. Through report generation and user interface, users can easily obtain, view and export analysis results to support related work such as water ecological monitoring.

[0216] In general, the present invention provides an automated chironomid larvae identification system by combining deep learning and image recognition technology, which can realize multi-focal plane image acquisition, image overlay processing, microstructure feature extraction and automatic species classification. The system has high recognition accuracy and processing efficiency, can effectively reduce manual intervention, improve the automated recognition and classification accuracy of chironomid larvae, and meet the needs of water ecological environment monitoring and biodiversity research. The system also provides real-time statistical analysis and report generation functions, providing comprehensive and timely data support for environmental monitoring.

[0217] Please refer to the attached Figure 8 The present invention also provides a method for identifying chironomid larvae based on deep learning and image recognition, comprising the following steps:

[0218] S1, an image acquisition step, obtaining multi-focal plane images of the chironomid larvae through an image acquisition module, wherein the image acquisition module includes a transmitted light source, an optical microscope and a high-resolution digital camera, and controls the digital camera to automatically acquire image sequences at different focal planes;

[0219] S2, an image stacking processing step, performing image stacking processing on the acquired multi-focal plane images, the image stacking processing comprising:

[0220] Each focal plane image is decomposed into multiple scale levels through multi-scale Laplacian pyramid decomposition;

[0221] The fusion weight is calculated according to the local gradient, contrast and motion compensation results of the image, and multiple focal plane images are fused to generate a full focal plane image;

[0222] S3, a feature extraction step, inputting the superimposed full-focus surface image into a deep learning recognition module, wherein the deep learning recognition module extracts features of the chironomid larvae based on a convolutional neural network, and the feature extraction process includes convolution, activation, and pooling operations;

[0223] S4, classification step, classifying the extracted features through the fully connected layer and Softmax classifier of the deep learning recognition module, and outputting the species label of the chironomid larvae and the corresponding classification confidence;

[0224] S5, a classification result storage step, storing the classification result together with the image number, timestamp, and detection position information into a database;

[0225] S6, statistical analysis step, statistical analysis of the classification results is performed through the automatic classification and statistical module, the number of different types of chironomid larvae and their time series changes are calculated, the species distribution and time trend graphs are generated, and a report is automatically generated.

[0226] The principles of the method steps of the present invention have been fully described in the aforementioned system implementation method. In order to avoid repeated description, they will not be repeated here.

[0227] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A chironomid larvae identification system based on deep learning and image recognition, characterized in that: include: An image acquisition module includes an optical microscope with a transmission light source, a microscope objective lens and a digital camera, wherein the digital camera is installed on the microscope and connected to a computer via a data transmission interface; the image acquisition module is controlled by computer software and automatically acquires multi-level images of chironomid larvae at different focal planes; An automated sample handling device, comprising a sample loading system and a three-axis motorized stage, wherein the sample loading system transports the chironomid larvae sample to an imaging position, and the three-axis motorized stage adjusts the XYZ axis position of the sample for automatic image acquisition in multiple fields of view and multiple focal planes; An image stacking processing module, wherein the image stacking processing module uses adaptive multi-scale image fusion to perform high-precision image stacking on the acquired images of different focal length planes to generate a full-focus plane image containing multiple depth information from bottom to top; The deep learning recognition module, based on the convolutional neural network model, automatically identifies the microstructural features of the chironomid larvae, including but not limited to the chin plate, inner labial comb, eyespots and caudal setae, by training and extracting features from the full-focus images generated after superposition, so as to achieve the classification of chironomid larvae of different subfamilies and genera; The module for automatic species classification and statistics automatically identifies and classifies the species of chironomid larvae by combining the output of the deep learning model, and provides statistical analysis. The results are stored in a database to support the needs of automated monitoring of biological monitoring and aquatic ecological environment.

2. The chironomid larvae identification system based on deep learning and image recognition according to claim 1, characterized in that: The sample loading system includes a peristaltic pump, which is connected to a sample liquid bottle through a sample feeding hose, and is connected to an imaging position of a three-axis electric stage through a sample delivery hose, and the imaging position of the three-axis electric stage is connected to a waste liquid barrel through a sample discharge hose.

3. The chironomid larvae identification system based on deep learning and image recognition according to claim 2, characterized in that: A sample containing groove is arranged in the middle of the three-axis electric stage, and the sample conveying hose and the sample discharging hose are both embedded in the three-axis electric stage, and the ends are connected in the sample containing groove.

4. The chironomid larvae identification system based on deep learning and image recognition according to claim 3, characterized in that: The upper surface of the three-axis electric stage is provided with a glass slide groove at the outer edge of the sample holding groove for placing glass slides and pressing the chironomid larvae for identification.

5. The chironomid larvae identification system based on deep learning and image recognition according to claim 1, characterized in that: The image overlay processing module comprises: A multi-focal plane image acquisition unit, used for acquiring an image sequence of the chironomid larvae from different focal planes, wherein the image sequence includes multiple images at different depths; The multi-scale image decomposition unit is used to perform Laplace pyramid decomposition on each image in the acquired image sequence to generate multiple image levels of different scales. The decomposition process is as follows: in, represents the Laplacian pyramid representation of the i-th focal plane image at the l-th scale, is the Gaussian pyramid representation of the image; The adaptive weighted fusion unit is used to calculate the fusion weight according to the local gradient, contrast and motion compensation results of the image. The weight calculation formula is: Among them, w i (x,y) is the weight of the i-th focal plane image, is the local gradient of the image, C i (x, y) is the local contrast, M i (x, y) is the image quality weight after motion compensation; The image fusion unit is used to fuse the images of each scale after adaptive weighted calculation. The fusion process is as follows: in, is the fused image at the lth scale; The image reconstruction unit is used to reconstruct the final full-focus image through an inverse Laplace pyramid process. The reconstruction formula is: Among them, I fused (x, y) is the final full-focal plane image after fusion, and L is the number of decomposed scales.

6. The chironomid larvae identification system based on deep learning and image recognition according to claim 5, characterized in that: The adaptive weighted fusion unit also includes an optical flow motion compensation unit, which is used to detect and correct the displacement caused by the sample movement in the liquid environment before the multi-focal plane image is superimposed. The motion compensation is calculated by the following formula: Among them, v(x,y) is the motion vector of the sample at time t, I(x,y,t) is the image grayscale value at time t, and the motion vector is used for correction before image superposition.

7. The chironomid larvae identification system based on deep learning and image recognition according to claim 6, characterized in that: The image reconstruction unit adopts the local feature enhancement technology in the process of reconstructing the full focal plane image, and enhances the details of the specific microstructure of the chironomid larvae through the Laplace operator. The reconstruction formula after enhancement is: Yo enhanced (x,y)=I fused (x,y)+λ·ΔI fused (x,y) Among them, ΔI fused (x, y) is the Laplacian operator result of the full focal plane image, and λ is the parameter that controls the enhancement intensity.

8. The chironomid larvae identification system based on deep learning and image recognition according to claim 1, characterized in that: The deep learning recognition module includes: An input preprocessing unit is used to perform normalization processing and data enhancement processing on the full-focus plane image generated by the image superposition processing module; The convolution feature extraction unit is used to extract high-level features from images through a convolutional neural network. The convolution process is: Among them, I normalized (x, y) is the normalized full-focal-plane image, w(i, j) is the weight of the convolution kernel, F(x, y) is the convolution result, and b is the bias; The activation unit is used to apply the ReLU activation function to the convolution result. The calculation formula of the activation function is: A(x,y)=max(0,F(x,y)) Among them, A(x,y) is the output after activation; The pooling unit is used to reduce the dimension and downsample the activated feature map through the maximum pooling operation. The pooling calculation formula is: Where P is the size of the pooling window; The fully connected layer unit is used to flatten the pooled feature map into a one-dimensional feature vector and map it to the target classification space. The calculation formula of the fully connected layer is: z=W·x+b Among them, W is the weight matrix of the fully connected layer, x is the input feature vector, and b is the bias; The Softmax classification unit is used to convert the output of the fully connected layer into the probability distribution of various types of chironomid larvae. The calculation formula of the Softmax classifier is: Among them, z j is the score of category j, N is the number of categories, and p(y=j|x) is the probability that the input image belongs to category j.

9. The chironomid larvae identification system based on deep learning and image recognition according to claim 1, characterized in that: It is characterized in that The automatic classification and statistics module includes: A classification result storage unit, used to receive the classification result of the deep learning recognition module, the classification result including the species label and confidence of the chironomid larvae, and store the classification result together with the corresponding image number, timestamp, and detection position metadata in a database; The confidence screening unit is used to screen according to the confidence in the classification result. When the classification confidence is lower than a preset threshold, the result is marked as an uncertain classification. The screening process is defined as: Among them, y i is the classification label, p i is the classification confidence, p threshold To preset a confidence threshold; The species statistics unit is used to perform statistical analysis on the stored classification results and calculate the number and proportion of different species of chironomid larvae. The statistical process is defined as: Among them, 1(y j =y i ) is the indicator function, when y j =y i Returns 1 when , otherwise returns 0, m is the number of samples; A time series analysis unit, used to perform time series analysis on the detection results of the species of chironomid larvae based on historical data, generate a trend graph of the number of different species of chironomid larvae changing over time, and dynamically monitor the detection data; The report generation unit is used to automatically generate a report based on the category statistics and time series analysis results. The report includes a category distribution diagram, a time change trend diagram, and supports exporting the report to PDF or Excel format.

10. A method for identifying chironomid larvae based on deep learning and image recognition, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: An image acquisition step, obtaining multi-focal plane images of the chironomid larvae through an image acquisition module, wherein the image acquisition module includes a transmitted light source, an optical microscope and a high-resolution digital camera, and controls the digital camera to automatically acquire image sequences at different focal planes; The image stacking processing step is to perform image stacking processing on the acquired multi-focal plane images, and the image stacking processing includes: Each focal plane image is decomposed into multiple scale levels through multi-scale Laplacian pyramid decomposition; The fusion weight is calculated according to the local gradient, contrast and motion compensation results of the image, and multiple focal plane images are fused to generate a full focal plane image; A feature extraction step, inputting the superimposed full-focus surface image into a deep learning recognition module, wherein the deep learning recognition module extracts features of the chironomid larvae based on a convolutional neural network, and the feature extraction process includes convolution, activation, and pooling operations; In the classification step, the extracted features are classified through the fully connected layer and Softmax classifier of the deep learning recognition module, and the species label of the midge larvae and the corresponding classification confidence are output; a classification result storage step, storing the classification result together with the image number, timestamp, and detection position information into a database; In the statistical analysis step, the classification results are statistically analyzed through the automatic species classification and statistical module, the number of different species of chironomid larvae and their time series changes are calculated, the species distribution and time trend graphs are generated, and a report is automatically generated.

Images