Insect egg recognition method, system, and storage medium

Abstract

This application discloses a method, system, and storage medium for identifying parasite eggs. The application relates to the field of medical testing technology. The parasite egg identification method includes: performing background suppression on an initial sample image to obtain a target sample image; the sample to be tested corresponding to the initial sample image includes a fecal slide sample or a urine filter membrane sample; inputting the target sample image into a domain-adaptive recognition model, which includes a backbone network and at least two branch networks, with the backbone network connected in parallel to at least two branch networks, and different sample types corresponding to different branch networks; the domain-adaptive recognition model selecting the corresponding branch network for inference based on the sample type of the target sample image to obtain the parasite egg identification result; and acquiring the parasite egg identification result output by the domain-adaptive recognition model, which includes location information, classification results, and confidence levels. This application can achieve the technical effect of improving the accuracy of parasite egg identification results.

Description

Technical Field

[0001] This application relates to the field of medical testing technology, and in particular to a method, system and storage medium for identifying parasite eggs. Background Technology

[0002] Currently, parasitic diseases are a serious global public health problem, and microscopic examination is commonly used to diagnose parasitic infections. Automated image recognition algorithms in related technologies are designed for clean slide backgrounds; however, different samples have different background interference characteristics. Using a uniform image processing or recognition path can affect the reliability of parasite egg identification, leading to low accuracy in egg identification results. Summary of the Invention

[0003] The main objective of this application is to provide a method, system, and storage medium for identifying insect eggs, aiming to solve the technical problem of low accuracy in insect egg identification results.

[0004] To achieve the above objectives, this application provides a method for identifying insect eggs, applied to a terminal device. The method for identifying insect eggs includes: Background suppression is performed on the acquired initial sample image to obtain the target sample image. The sample to be tested corresponding to the initial sample image includes fecal slide sample or urine filter membrane sample. The target sample image is input into the domain-adaptive recognition model, which includes a backbone network and at least two branch networks. The backbone network is connected in parallel to at least two branch networks, and different sample types correspond to different branch networks. The domain-adaptive recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg recognition result. Obtain the insect egg identification results output by the domain adaptation recognition model. The insect egg identification results include location information, classification results, and confidence scores.

[0005] In one embodiment, the step of performing background suppression on the acquired initial sample image to obtain the target sample image includes: When the sample to be tested is a urine filter membrane sample, the initial sample image is transformed from the spatial domain to the frequency domain to obtain the first frequency domain image. The initial sample image is an image with a periodic grid texture. A band-stop filter is set based on the grid texture features of the filter membrane in the initial sample image. The periodic grid background noise in the initial sample image is filtered out based on the band-stop filter to obtain the second frequency domain image. Perform an inverse Fourier transform on the second frequency domain image to convert it from the frequency domain to the spatial domain, thus obtaining the target sample image after background suppression.

[0006] In one embodiment, before the step of performing background suppression on the acquired initial sample image to obtain the target sample image, the method further includes: Place the sample to be tested on the stage, perform a preview scan on the sample to be tested, and acquire a preview image; The preview image is input into a preset classification network for recognition to obtain the sample type of the sample to be tested; The scanning strategy for the sample to be examined is determined based on the sample type. The scanning strategy includes the imaging mode and objective magnification. The sample to be inspected is scanned based on a scanning strategy to obtain an initial sample image.

[0007] In one embodiment, after the step of obtaining the insect egg identification result output by the domain adaptation identification model, the method further includes: If the confidence level is greater than or equal to the confidence level threshold, and the ratio of the highest category probability to the second highest category probability in the classification result is higher than the preset ratio threshold, a detection report is generated based on the location information and the classification result; the confidence level is the probability of the presence of insect eggs, and the category probability is the probability corresponding to the classification result; The target sample image is labeled as a regular image.

[0008] In one embodiment, after obtaining the insect egg identification result output by the domain adaptation identification model, the method further includes: When at least one preset condition is met, the target sample image is marked as a difficult image; Among them, meeting the preset conditions includes: The confidence level is less than the confidence threshold; When the overlapping area of ​​at least two detection boxes that mark the location of insect eggs is greater than a preset area threshold, and the highest classification probability of each detection box is higher than a preset classification threshold, the classification results corresponding to the highest classification probabilities of each detection box are inconsistent; the detection box is the bounding box marked by the domain adaptation recognition model in the target sample image to locate the location of insect eggs. The number of detection boxes output by the domain adaptation recognition model is 0, and the texture complexity of the target sample image is greater than the complexity threshold.

[0009] In one embodiment, the method further includes: Upload the detection report, keyframe index, and regular image to the cloud platform. The keyframe index is the image region number of the regular image where the insect eggs are located. Upload the problematic images and corresponding insect egg identification results to the cloud platform; Receive differential model parameters sent by the cloud platform, and update the parameters of the domain adaptation recognition model based on the differential model parameters.

[0010] Furthermore, to achieve the above objectives, this application also provides a method for identifying insect eggs, applied to a cloud platform. The method for identifying insect eggs includes: The system receives detection reports, keyframe indexes, and corresponding regular images sent by the receiving terminal device, as well as difficult images and corresponding insect egg identification results. The keyframe index is the image region number of the regular image where the insect egg is located. The detection report, keyframe index, and corresponding regular image are used as regular samples; Obtain manually reviewed and corrected insect egg identification results from problematic images, and integrate the corrected insect egg identification results with the problematic images to obtain problematic samples; Incremental optimization is performed on the domain adaptation recognition model stored in the cloud platform based on regular samples and difficult samples to obtain the optimized model parameters; Determine the differential model parameters between the optimized model parameters and the domain adaptation recognition model parameters, and send the differential model parameters to the terminal device.

[0011] In one embodiment, the step of incrementally optimizing the domain adaptation recognition model stored in the cloud platform based on regular samples and difficult samples to obtain the optimized model parameters includes: A mixed dataset is constructed by combining difficult samples with regular samples, and the domain-adaptive recognition model is used as the baseline model. The baseline model was incrementally trained using a mixed dataset, and the branch networks of the baseline model were trained using sample data of the corresponding sample types to obtain the optimized model parameters.

[0012] Furthermore, to achieve the above objectives, this application also provides an insect egg identification system. The insect egg identification system adopts a cloud-edge-device collaborative architecture. The insect egg identification system includes a terminal device and a cloud platform. The terminal device includes: a microscopic imaging module, an electric stage, an edge computing module, a battery module, a communication module, a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The computer program is configured to implement the steps of the insect egg identification method described above. The cloud platform includes: a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The computer program is configured to implement the steps of the insect egg identification method described above.

[0013] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described insect egg identification method.

[0014] This application provides a method for insect egg identification. The method involves obtaining a target sample image by performing background suppression on an initial sample image. The initial sample image corresponds to a sample to be inspected, including fecal slide samples or urine filter membrane samples. The target sample image is input into a domain-adaptive recognition model, which includes a backbone network and at least two branch networks. The backbone network connects to at least two branch networks in parallel, with different branch networks corresponding to different sample types. The domain-adaptive recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg identification result. The insect egg identification result output by the domain-adaptive recognition model is then obtained, including location information, classification results, and confidence scores. By accurately filtering out background noise, a clear input image with target features is provided. Simultaneously, the domain-adaptive recognition model ensures adaptive identification of two different types of samples through dedicated branch networks, further filtering to obtain highly reliable results, thus improving the accuracy of insect egg identification. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an embodiment of the insect egg identification method of this application. Figure 2 A schematic diagram illustrating the construction and working principle of the domain adaptive recognition model provided in Embodiment 1 of the insect egg recognition method of this application; Figure 3 This is a schematic diagram of the microscopic imaging recognition display interface provided in Embodiment 1 of the insect egg identification method of this application; Figure 4 This is a low-power scanning path planning diagram provided in Embodiment 3 of the insect egg identification method of this application; Figure 5 This is a flowchart illustrating Embodiment Seven of the insect egg identification method of this application; Figure 6 This is a schematic diagram of the cloud platform expert review interface provided in Embodiment 7 of the insect egg identification method of this application; Figure 7 A simplified flowchart illustrating the insect egg identification method of this application; Figure 8 A schematic diagram of the structure of the terminal device provided in this application; Figure 9 A schematic diagram of the overall architecture of the insect egg identification system provided in this application; Figure 10 This is a schematic diagram of the insect egg identification system in an embodiment of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] Currently, parasitic diseases remain a serious global public health problem. The World Health Organization (WHO) recently estimated that approximately 1.2 billion people worldwide are infected with at least one neglected tropical parasitic disease (NTD), accounting for nearly 15% of the global population. Among these, malaria, despite significant progress in control efforts in recent years, still reported approximately 282 million cases globally in 2024, with annual deaths remaining around 610,000. Schistosomiasis is prevalent in 79 countries, infecting over 250 million people. Soil-transmitted worms are particularly prevalent among children in impoverished areas, leading to malnutrition, developmental delays, and cognitive impairment, creating a vicious cycle of infection and poverty. Furthermore, global travel and migration have resulted in a year-on-year increase in imported parasitic cases, placing continuous pressure on healthcare systems in various countries.

[0022] Traditional parasitological examination methods primarily rely on morphological examination using optical microscopy. Microscopic examination, considered the "gold standard," typically involves preparing smears of fecal, urine, blood, or tissue samples, followed by manual microscopic observation of the morphology and count of parasite eggs. It offers advantages such as low cost, minimal equipment requirements, and direct identification of parasite species and infection intensity. However, its limitations are equally prominent: results are highly dependent on the experience of the examiner, the procedure is time-consuming and inefficient, and diagnostic accuracy is significantly affected by sample quality. These limitations make traditional methods insufficient for precise prevention and control, necessitating technological innovation.

[0023] In recent years, image recognition technology based on artificial intelligence algorithms has become a research hotspot. However, most existing image recognition technologies are designed for clean glass slide backgrounds, while different samples to be inspected have different background interference characteristics. Using a uniform image processing or recognition path will affect the reliability of insect egg identification, leading to a decrease in detection accuracy. Therefore, it is necessary to develop models with adaptive background removal, cross-domain transfer learning, and multi-feature fusion capabilities to adapt to the highly heterogeneous real world.

[0024] The main solution of this application is as follows: Target sample images are obtained by performing background suppression on the acquired initial sample images. The samples to be tested corresponding to the initial sample images include fecal slide samples or urine filter membrane samples. The target sample images are input into a domain-adaptive recognition model, which includes a backbone network and at least two branch networks. The backbone network connects to at least two branch networks in parallel, with different branch networks corresponding to different sample types. The domain-adaptive recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg recognition result. The insect egg recognition result output by the domain-adaptive recognition model is obtained, including location information, classification results, and confidence scores. By accurately filtering out background noise, a clear input image with target features is provided. Simultaneously, the domain-adaptive recognition model ensures adaptive recognition of two different types of samples through dedicated branch networks, and further filters to obtain highly reliable results, thus improving the technical effect of improving the accuracy of insect egg recognition results.

[0025] It should be noted that the executing entity in this embodiment can be an insect egg identification system, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an insect egg identification system capable of performing the above functions. This embodiment does not specifically limit this. The following uses an insect egg identification system as the executing entity to describe this embodiment and the following embodiments.

[0026] Based on this, Embodiment 1 of this application proposes a method for identifying insect eggs. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the insect egg identification method of this application, applied to a terminal device. The insect egg identification method includes steps S10 to S30: Step S10: Background suppression is performed on the acquired initial sample image to obtain the target sample image. The sample to be tested corresponding to the initial sample image includes fecal slide sample or urine filter membrane sample.

[0027] Among them, fecal slide samples refer to thin smears made by uniformly spreading fecal samples onto glass slides after specific treatments, such as filtration and staining. These are routine microscopic examination samples for diagnosing intestinal parasite infections. Urine filter membrane samples refer to urine samples that have been filtered and enriched through a filter membrane with precise micropores, trapping any possible parasite eggs on the membrane, which is then placed on a glass slide.

[0028] As an optional implementation, when the sample to be tested is a urine filter membrane sample, the initial sample image is subjected to background suppression using a Fourier transform algorithm to obtain the target sample image.

[0029] Specifically, when the sample to be tested is a urine filter membrane sample, an image of the urine filter membrane sample acquired by the microscopic imaging module is obtained. This initial sample image contains background texture noise due to the periodic grid structure of the filter membrane itself. The Fast Fourier Transform (FFT) algorithm is used to convert the image to the frequency domain. At this point, the grid texture appears as a series of regular high-energy bright spots in the spectral image. The locations of the bright spots are pre-determined or adaptively determined based on the grid texture characteristics of the filter membrane, and a corresponding band-stop filter is designed for energy attenuation. Then, the processed frequency domain image is converted back to the spatial domain through inverse Fourier transform to obtain the target sample image. The inherent periodic grid noise of the filter membrane is accurately eliminated through frequency domain band-stop filtering, thereby improving the recognition accuracy of subsequent models on the filter membrane.

[0030] As another optional implementation, when the sample to be tested is a fecal slide sample, the initial sample image is preprocessed using image morphology operations or threshold segmentation strategies to obtain the target sample image after background suppression.

[0031] Specifically, when the sample to be tested is a fecal slide, morphological operations such as erosion, dilation, opening, and closing are performed on this binary or grayscale image to eliminate noise points smaller than the structuring element, fill in voids inside the target, or smooth the target edge to obtain a cleaner and more complete candidate target region; or a threshold segmentation strategy is used, which automatically calculates and selects one or more grayscale thresholds based on the image's grayscale histogram to binarize the image into the foreground, i.e., the target of suspected insect eggs and the background, thereby achieving separation of the target and background, enhancing image contrast and preparing for subsequent identification, avoiding unnecessary complex calculations, and optimizing the overall processing efficiency.

[0032] Step S20: Input the target sample image into the domain adaptation recognition model. The domain adaptation recognition model includes a backbone network and at least two branch networks. The backbone network is connected in parallel to at least two branch networks. Different sample types correspond to different branch networks. The domain adaptation recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg recognition result.

[0033] The domain-adaptive recognition model refers to a recognition model designed to handle multiple types of samples. It comprises a shared backbone network and multiple specialized branch networks fine-tuned for different sample domains. The backbone network is a deep neural network pre-trained on a large-scale dataset, serving as a shared feature extractor responsible for extracting common low-level features across sample types, such as edges, textures, and morphology. The branch networks are specialized network parts attached to the backbone network, fine-tuned using data from specific sample types. Each branch network specifically learns the unique high-level feature distribution of its particular sample domain. The domain-adaptive recognition model can automatically route to the corresponding branch network for inference based on the type of the input sample to be detected, outputting a complete insect egg recognition result. This domain-adaptive recognition model, including a backbone network and at least two branch networks, fully utilizes the common knowledge of multi-source data through a parameter-sharing mechanism and avoids feature distribution conflicts between different sample types through branch isolation, effectively mitigating the domain shift problem. While maintaining the deployment efficiency of a single model, it achieves adaptive processing of heterogeneous samples, improving the recognition accuracy in multi-source sample scenarios.

[0034] As an optional implementation, the target sample image is input into the domain-adaptive recognition model. The domain-adaptive recognition model selects the corresponding branch network for inference according to the sample type to obtain the recognition result of each insect egg. Among them, fecal slide samples are routed to the fecal recognition model, and urine filter membrane samples are routed to the urine recognition model.

[0035] Specifically, please refer to Figure 2 , Figure 2 This diagram illustrates the construction and working principle of the domain-adaptive recognition model provided in Embodiment 1 of the insect egg identification method of this application. The target sample image is input into the domain-adaptive recognition model. Based on the determined sample type, the image is automatically routed to the corresponding dedicated branch network for inference calculation. If the sample type is a fecal slide, it is routed to the fecal recognition model; if the sample type is a urine filter membrane, it is routed to the urine recognition model. After inference, the model outputs a structured result including location information, classification result, and confidence score for each suspected insect egg target in the image. Please refer to [reference needed]. Figure 3 , Figure 3 This is a schematic diagram of the microscopic imaging recognition display interface provided in Embodiment 1 of the insect egg recognition method of this application. By extracting common features through a shared backbone network and routing to dedicated branch networks according to sample type, adaptive and high-precision insect egg recognition for two significantly different types of samples is achieved.

[0036] As an optional implementation, the domain-adaptive recognition model adopts a dynamic routing architecture. A learnable routing gating unit is set between the backbone network and each branch network. The routing gating unit automatically calculates and outputs a weight vector based on the features of the input image. The domain-adaptive recognition model aggregates the weighted outputs of each branch to obtain the insect egg recognition result.

[0037] Specifically, during the forward propagation of the model, image features, after being extracted by the backbone network, first enter the routing gating unit. The routing gating unit calculates an attention score for each branch network. For branches that highly match the input image domain, such as the urine branch corresponding to a urine filter sample image, the attention score is close to 1, and the features are almost completely passed; for branches that do not match, the score is close to 0, and their feature contribution is ignored. The features extracted by the backbone network are fed in parallel into all dedicated branch networks, and each branch performs inference independently, generating the corresponding preliminary recognition result. The domain-adaptive recognition model aggregates the outputs of each branch, that is, multiplies the preliminary result of each branch by its corresponding routing attention score to obtain the egg recognition result. For example, the attention score of the urine branch is 0.95, and that of the feces branch is 0.05, meaning the output result of the urine branch accounts for 95% of the weight during aggregation, while the result of the feces branch accounts for only 5%. Through the routing gating unit, feature-based adaptive branch selection is achieved, enabling the model to integrate professional knowledge from different domains with optimal weights, improving the recognition accuracy and generalization ability for mixed or fuzzy domain samples.

[0038] Step S30: Obtain the insect egg identification results output by the domain adaptation recognition model. The insect egg identification results include location information, classification results, and confidence scores.

[0039] The location information refers to the pixel-level bounding box coordinates that identify the specific location of each insect egg instance in the image. The classification result refers to the probability distribution of each detected insect egg instance belonging to each preset species. The confidence score refers to the certainty score of whether a real insect egg exists within each detection box.

[0040] As an optional implementation, after the domain adaptation recognition model completes the inference calculation of the target sample image, the positioning information, classification result and confidence level corresponding to each detection instance are parsed and extracted from the structured data output by the model in a preset format.

[0041] Specifically, the pixel coordinates of the detection boxes are read from the structured data output by the model. For example, four pixel coordinate values ​​representing the top-left and bottom-right corners of a rectangle are used to obtain the bounding box coordinates that identify the specific location of the insect egg instance in the image, i.e., the localization information. The classification vector corresponding to the detection box is read. This classification vector is a probability array with a length equal to the preset number of insect egg species, where each element represents the probability that the target belongs to the corresponding insect egg species. This yields the probability distribution of each detected insect egg instance belonging to each preset species, i.e., the classification result. The scalar value attached to the detection box, between 0 and 1, is read. This value is the model's deterministic measure of the presence of real insect eggs within the box, yielding a deterministic score for the presence of real insect eggs within each detection box, i.e., the confidence score. By parsing and extracting the structured results output by the model, a precise data foundation is provided for subsequent result judgment, report generation, and cloud collaboration.

[0042] As another optional implementation, the insect egg identification result output by the domain adaptation recognition model is obtained. The insect egg identification result also includes the number of insect eggs. The insect egg density is calculated based on the number of insect eggs and the current field of view area, and the objective lens magnification is adaptively adjusted based on the insect egg density.

[0043] Specifically, the insect egg density is calculated as the number of insect eggs in the current field of view divided by the area of ​​the current field of view. When the insect egg density is less than the first density threshold, the control program automatically switches to a high-power objective lens for fine scanning to improve the detection capability of small or sparsely distributed insect eggs. When the insect egg density is greater than the second density threshold, the program automatically switches to a low-power objective lens to expand the coverage of a single field of view and avoid repeated movement and wasted time in high-density areas due to a small field of view.

[0044] This embodiment first obtains the target sample image by performing background suppression on the acquired initial sample image. The initial sample image corresponds to samples including fecal slide samples or urine filter membrane samples. The target sample image is then input into a domain-adaptive recognition model, which includes a backbone network and at least two branch networks. The backbone network connects to at least two branch networks in parallel, with different branch networks corresponding to different sample types. The domain-adaptive recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg recognition result. The insect egg recognition result output by the domain-adaptive recognition model is then obtained, including location information, classification results, and confidence scores. By accurately filtering out background noise, a clear input image with target features is provided. At the same time, the domain-adaptive recognition model ensures adaptive recognition of two different types of samples through dedicated branch networks, and further filters to obtain highly reliable results, thus improving the technical effect of improving the accuracy of insect egg recognition results.

[0045] Based on Embodiment 1, in Embodiment 2 of this application, the content that is the same as or similar to that in Embodiment 1 can be referred to the above description, and will not be repeated hereafter. Based on this, the steps of performing background suppression on the acquired initial sample image to obtain the target sample image include: Step S11: When the sample to be tested is a urine filter membrane sample, the initial sample image is converted from the spatial domain to the frequency domain to obtain the first frequency domain image. The initial sample image is an image with periodic grid texture.

[0046] In this embodiment, the initial sample image refers to the digital image of a urine filter membrane sample carrying parasite eggs, acquired through a microscopic imaging module. The spatial domain refers to the image directly represented by pixels and their grayscale or color values, reflecting the spatial distribution of the target. The frequency domain refers to the representation of the image after Fourier transform, reflecting the intensity distribution of different frequency components in the image, such as smooth backgrounds, edges, and periodic textures. Frequency refers to the rate of change in brightness in the image; low-frequency components correspond to gently changing, large areas in the image, such as a uniform background or the interior of a large target; high-frequency components correspond to rapidly changing, sharp parts in the image, such as object edges, textures, and noise. The first frequency domain image refers to the visual representation of the initial sample image in the frequency domain, i.e., a spectrum, where periodic textures are presented as bright frequency points. Periodic grid textures refer to the regularly repeating grid-like background pattern formed by the physical structure of the standard filter membrane itself in the microscopic image, which is a strong form of interference noise.

[0047] As an optional implementation, an initial sample image acquired by the microscopic imaging module is obtained. This initial sample image has a periodic grid texture generated by the physical structure of the filter membrane itself. In the edge computing module, an optimized Fast Fourier Transform (FFT) library function is called. This library function is optimized at the low level for the instruction set and memory hierarchy of the embedded AI computing chip, such as using single-instruction multiple-data-stream instruction parallel processing of pixel data, or employing a block computing strategy to efficiently utilize the cache. The input initial sample image, i.e., the pixel matrix in the spatial domain, is used as data. Its two-dimensional discrete Fourier transform is calculated to convert it into a frequency domain representation, resulting in a frequency domain image in the form of a complex matrix, i.e., the first frequency domain image. In this frequency domain image, the periodic grid texture appears as a series of regularly arranged high-energy bright spots. By employing a Fast Fourier Transform algorithm optimized for embedded hardware, efficient real-time conversion of the filter membrane image from the spatial domain to the frequency domain is achieved.

[0048] Step S12: Set a band-stop filter based on the grid texture features of the filter membrane in the initial sample image, and filter out the periodic grid background noise in the initial sample image based on the band-stop filter to obtain the second frequency domain image.

[0049] In this embodiment, the mesh texture feature of the filter membrane refers to the physical dimensions of the mesh texture of the filter membrane. For example, the physical dimensions of a standard filter membrane include an pore size of 12-35 μm and a diameter of 13 mm. The physical parameters of the filter membrane determine the spatial period of the mesh in the image, i.e., the fixed distance of the texture interval. A band-stop filter is a frequency selection tool used to precisely block one or more signal components within a specific frequency range in a frequency domain image, while allowing other frequency components to pass through.

[0050] As an optional implementation, the spatial period of the filter membrane's mesh texture features, such as its pore size, is calculated in the image; that is, the pixel distance between two adjacent mesh points. Based on the image size and the sampling theorem, this spatial period is converted into corresponding frequency coordinates in the frequency domain image. Using these coordinates as the center, a band-stop filter with a stopband radius of R is set, effectively zeroing out the frequency components within this circular region in the frequency domain. The designed band-stop filter is multiplied by the frequency domain image to filter out specific frequency components representing mesh noise, resulting in a second frequency domain image. By transforming the known mesh texture features of the filter membrane into a precisely located band-stop filter in the frequency domain, targeted filtering of periodic mesh noise is achieved, thereby improving the model's accuracy in recognizing mesh noise in the filter membrane image.

[0051] Step S13: Perform an inverse Fourier transform on the second frequency domain image to convert the second frequency domain image from the frequency domain to the spatial domain, and obtain the target sample image after background suppression.

[0052] As an optional implementation, a low-level optimized inverse Fourier transform library function, matching the forward transform, is invoked. This function takes a second frequency domain image, i.e., a complex matrix after band-stop filtering, as input and converts it into a real pixel matrix in the spatial domain through inverse transform calculation, thus obtaining the target sample image. By reconstructing the image from a frequency domain representation after eliminating specific noise frequency components into a spatial domain image that can be directly observed by the human eye and processed by subsequent models, a clear and direct input of target features is provided for the subsequent model's egg identification, solving the identification bottleneck caused by filter background interference.

[0053] In this embodiment, frequency domain band-stop filtering is used to eliminate the inherent periodic grid noise of the filter membrane, providing a high signal-to-noise ratio input image for subsequent model recognition and improving the recognition accuracy on the filter membrane sample.

[0054] Based on any of the above embodiments of this application, Embodiment 3 of this application proposes a method for identifying insect eggs, which can be referred to the above description and will not be repeated hereafter. Based on this, before performing a Fourier transform on the acquired filter membrane image, the method includes: Step S101: Place the sample to be inspected on the stage, perform a preview scan on the sample to be inspected, and acquire a preview image.

[0055] As one implementation method, after placing the sample to be examined onto the stage, the stage is moved, and one or a few preview images covering the overall appearance of the sample are quickly captured using a low-power objective lens. By quickly previewing the sample, low-resolution images reflecting the macroscopic features of the sample can be obtained with minimal time cost, avoiding the efficiency loss that may result from directly performing a full-area high-resolution scan.

[0056] Step S102: Input the preview image into the preset classification network for recognition to obtain the sample type of the sample to be tested.

[0057] As one implementation method, the acquired preview image is input into a pre-trained lightweight classification neural network for real-time inference. This network can quickly determine whether the current sample type is a fecal slide sample or a urine filter membrane sample based on the macroscopic features of the image. This intelligent identification of the sample type provides a basis for subsequent differentiated processing decisions.

[0058] Step S103: Determine the scanning strategy for the sample to be inspected based on the sample type. The scanning strategy includes the imaging mode and objective lens magnification.

[0059] Step S104: Scan the sample to be inspected based on the scanning strategy to obtain an initial sample image.

[0060] As one implementation method, the operating parameters of all subsequent hardware and software modules are dynamically configured based on the sample type. For example, if the sample is identified as a urine filter membrane, a polarized light illumination mode is used, and a 4X objective lens is selected to scan and obtain a polarized image of the area to be detected. Based on the birefringence response intensity of the area to be detected in the polarized image, the crystal interference area is marked, eliminated, or its recognition priority as a candidate target for insect eggs is reduced, thereby effectively suppressing the interference of crystal particles in urine on the identification of insect eggs. Polarized light can distinguish between insect eggs and crystals by utilizing the characteristic that insect eggs have no birefringence while crystals have strong birefringence. Please refer to [reference needed]. Figure 4 , Figure 4 The scanning path planning diagram under low magnification provided in Embodiment 3 of the insect egg identification method of this application is as follows: The control program first calculates the physical coverage of a single field of view based on the current magnification of the objective lens and the camera sensor parameters. Based on the known grid texture characteristics of the filter membrane, the scanning step size is set to an integer multiple of the grid period, and the entire area is traversed in a serpentine path to generate a sequence of motion commands containing the coordinates of each field of view, thereby achieving optimal full-coverage scanning. If the sample is identified as a fecal slide, the standard bright-field illumination mode is activated, and a suitable objective magnification, such as a 10X objective lens, is selected for fully automatic scanning of the slide. By assigning appropriate scanning strategies to samples with different characteristics, the overall processing efficiency is optimized.

[0061] In this embodiment, by quickly previewing and configuring the sample type, the appropriate scanning strategy can be accurately allocated for subsequent scanning and image processing based on the real-time judgment of the sample type, ensuring the detection accuracy for different samples to be inspected and improving the processing efficiency of insect egg identification.

[0062] Based on any of the above embodiments of this application, Embodiment 4 of this application proposes a method for insect egg identification, which can be referred to the above description and will not be repeated hereafter. Following the step of obtaining the insect egg identification result output by the domain adaptation identification model, the method further includes: Step S31: If the confidence level is greater than or equal to the confidence level threshold, and the ratio of the highest category probability to the second highest category probability in the classification result is higher than the preset ratio threshold, a detection report is generated based on the location information and the classification result; the confidence level is the probability of the presence of insect eggs, and the category probability is the probability corresponding to the classification result.

[0063] Step S32: Mark the target sample image as a regular image.

[0064] As one implementation method, each detection box is traversed. A reliable result is defined as one where both the confidence level is greater than or equal to a preset confidence threshold (e.g., 0.85) and the ratio of the highest class probability to the second-highest class probability in the classification results is higher than a preset ratio threshold. The location information and classification result are then adopted. All adopted reliable results in the target sample image are summarized, and a structured detection report containing the type, quantity, coordinates, and confidence level of the insect eggs is automatically generated. The target sample image is then marked as a regular image. Through dual threshold judgments of confidence level and classification clarity, automated quality filtering of the recognition results is achieved, ensuring that highly reliable results are adopted and reports are generated.

[0065] In this embodiment, by using the dedicated branch network architecture of the domain-adaptive recognition model and the dual threshold determination based on confidence and classification clarity, it is possible to adaptively and accurately process two significantly different test samples, fecal slides and urine filter membranes, thereby improving the automation level of the detection process and the reliability of the output results.

[0066] Based on any of the above embodiments of this application, Embodiment 5 of this application proposes a method for insect egg identification, which can be referred to the above description and will not be repeated hereafter. In addition, after obtaining the insect egg identification result output by the domain adaptation identification model, the method further includes: Step S33: When at least one preset condition is met, the target sample image is marked as a difficult image.

[0067] Among them, meeting the preset conditions includes: The confidence level is less than the confidence threshold.

[0068] When the overlapping area of ​​at least two detection boxes that mark the location of insect eggs is greater than a preset area threshold, and the highest classification probability of each detection box is higher than a preset classification threshold, the classification results corresponding to the highest classification probabilities of each detection box are inconsistent; the detection box is the bounding box marked by the domain adaptation recognition model in the target sample image to locate the location of insect eggs.

[0069] The number of detection boxes output by the domain adaptation recognition model is 0, and the texture complexity of the target sample image is greater than the complexity threshold.

[0070] As one implementation method, if the confidence level of any detection box is lower than a preset confidence threshold, the image is marked as a difficult image; or if there are two or more detection boxes indicating the location of insect eggs, the overlapping area of ​​their bounding boxes exceeds a preset area threshold, and their respective highest classification probabilities are all higher than a preset classification probability threshold, but the classification results corresponding to the highest probabilities are inconsistent, indicating a classification conflict within the model itself, the image is marked as a difficult image; or if the model does not output any detection boxes, i.e., it is determined that there are no insect eggs, but the texture complexity of the image calculated by the image analysis algorithm is higher than a preset complexity threshold, indicating the presence of complex features and potentially missed detection areas, the image is marked as a difficult image. By setting quantifiable automatic triggering rules, all potential uncertainties and high-risk situations in the model's inference results can be proactively identified.

[0071] In this embodiment, by identifying suspicious targets with low confidence, areas where the model itself makes conflicting judgments, and complex backgrounds that may be missed, and marking these images as difficult images, it is possible to effectively prevent low-quality or erroneous identification results from being included in the report. At the same time, it realizes the automatic screening and collection of difficult cases, providing a clear target for subsequent cloud-based expert review.

[0072] Based on any of the above embodiments of this application, Embodiment Six of this application proposes a method for identifying insect eggs, which can be referred to the above description and will not be repeated hereafter. In addition, the method further includes: Step S34: Upload the detection report, keyframe index, and regular image to the cloud platform. The keyframe index is the image region number of the regular image where the insect eggs are located.

[0073] As one implementation method, a lightweight image processing program is invoked to reduce the size and compress the quality of the original regular image, generating a low-resolution thumbnail with a small data size. A detection report containing the type, quantity, location, and confidence level of the insect eggs, a keyframe index for locating the specific image region where the eggs are located, and the corresponding low-resolution thumbnail of the regular image are packaged and uploaded to the cloud platform. By transmitting the detection conclusions and key evidence index information of the tested samples to the cloud, the cloud's need for recording and statistically analyzing detection results is met, while saving network bandwidth and perfectly adapting to weak network conditions in outdoor or resource-constrained environments.

[0074] Step S35: Upload the problematic images and the corresponding insect egg identification results to the cloud platform.

[0075] As one implementation method, high-resolution original images of problematic images and their corresponding insect egg identification results are uploaded to a cloud platform. The detection results include location information, classification results, and confidence levels. By submitting image cases with uncertainty or high risk of missed detection in the model, along with their preliminary analysis results, to the cloud platform for expert review, it is ensured that experts can efficiently focus on truly difficult problems.

[0076] Step S36: Receive the differential model parameters sent by the cloud platform, and update the parameters of the domain adaptation recognition model based on the differential model parameters.

[0077] As one implementation method, after model optimization and updating on the cloud platform, the system receives differential model parameters from the cloud platform, i.e., the difference data packet between the weights of the old and new models, and updates the domain adaptation recognition model based on this difference data packet. By transmitting only the difference portion of the model parameters, the network traffic and update time required for model updates are reduced.

[0078] In this embodiment, lightweight data is uploaded for routine results, while original images are uploaded for complex cases. With limited network bandwidth, this achieves cloud synchronization and recording of all detection results while accurately pushing uncertain samples to experts for review. Furthermore, by receiving and applying differential model parameters from the cloud, new features learned from global data and expert knowledge can be integrated at low cost, enabling continuous adaptive optimization of recognition capabilities.

[0079] Based on the above embodiments of this application, Embodiment Seven of this application proposes a method for identifying insect eggs. Please refer to... Figure 5 , Figure 5 This is a flowchart illustrating Embodiment Seven of the insect egg identification method of this application, applied to a cloud platform. The insect egg identification method includes: Step S40: Receive the detection report, keyframe index and corresponding regular image sent by the terminal device, as well as the difficult image and corresponding insect egg identification result. The keyframe index is the image region number of the regular image where the insect egg is located.

[0080] In one implementation, the cloud platform receives two types of data uploaded from terminal devices: lightweight data including detection reports for all samples to be inspected for regular images, keyframe indexes, and corresponding thumbnails of the regular images; and high-resolution original images and preliminary detection results for problematic images. By receiving the lightweight data for regular images and the detailed data for problematic images, a complete data source is provided for subsequent expert review and model optimization.

[0081] Step S50: The detection report, keyframe index, and corresponding regular image are used as regular samples.

[0082] As one implementation method, the keyframe indexes corresponding to the high-confidence images identified as regular images are extracted from the detection report, and associated with stored regular image thumbnails. The detection report, keyframe indexes, and regular image thumbnails are then packaged together as a regular sample. By packaging the core conclusions with lightweight images, network bandwidth utilization efficiency is optimized while ensuring complete detection information is obtained.

[0083] Step S60: Obtain the manually verified and corrected insect egg identification results of the problematic images, and integrate the corrected insect egg identification results with the problematic images to obtain problematic samples.

[0084] As one implementation method, experts use a dedicated review interface provided by the cloud platform, for example, please refer to... Figure 6 , Figure 6 This is a schematic diagram of the cloud platform expert review interface provided in Embodiment 7 of the insect egg identification method of this application. The review page includes three sub-interfaces: sample information, original image, and review operation. Experts can determine, correct, or re-label the recognition results of difficult images based on the sample information containing the model's preliminary recognition results and the high-resolution original image, obtaining the corrected recognition results. The difficult images and the corrected recognition results are then integrated into difficult samples. Through expert experience, edge cases, ambiguous targets, and suspected new features that are difficult for the model to handle are transformed into high-quality training data, which can solve the cognitive blind spots that cannot be handled by model self-learning.

[0085] Step S70: Based on regular samples and difficult samples, perform incremental optimization on the domain adaptation recognition model stored in the cloud platform to obtain the optimized model parameters.

[0086] As one implementation method, a hybrid dataset is constructed by aggregating regular samples with challenging samples that have been reviewed and corrected by experts. The challenging samples include challenging images and their accurate annotations, used to teach the model to recognize previously misleading, ambiguous, or unseen new features. Regular samples include regular images and their detection results, used to reinforce the model's existing recognition capabilities and prevent the model from forgetting old knowledge on new data. The hybrid dataset is then used to incrementally optimize the domain-adaptive recognition model; that is, based on the existing model knowledge, the model weights are further fine-tuned and optimized using new data to obtain optimized model parameters. By incrementally optimizing the domain-adaptive recognition model based on both regular and challenging samples, it can learn new features from challenging cases while maintaining recognition accuracy for regular samples.

[0087] Step S80: Determine the differential model parameters between the optimized model parameters and the domain adaptation recognition model parameters, and send the differential model parameters to the terminal device.

[0088] As one implementation method, after incremental training of the model is completed on the cloud platform, the optimized weight parameters are compared with the weight parameters of the currently used domain adaptation recognition model, and the difference between the two is calculated, i.e., the differential model parameters. The cloud platform then sends this differential parameter packet to the terminal device. By transmitting only the changed parameters of the model, the network traffic and transmission time required for model updates are reduced.

[0089] In this embodiment, by integrating regular samples with difficult samples that have been reviewed and corrected by experts, the domain adaptation recognition model is incrementally optimized in a targeted manner. This allows the model to continuously learn from real-world usage scenarios and expert experience, driving the continuous adaptive evolution of model performance and improving detection capabilities.

[0090] Based on any of the above embodiments of this application, Embodiment Eight of this application proposes a method for insect egg identification, which can be referred to the above description and will not be repeated hereafter. Based on this, the steps of incrementally optimizing the domain adaptation identification model stored in the cloud platform based on regular samples and difficult samples to obtain the optimized model parameters include: Step S71: Combine the difficult samples with the regular samples to form a mixed dataset, and use the domain adaptation recognition model as the baseline model.

[0091] As one implementation method, challenging samples are combined with regular samples to form a hybrid dataset for this training. The cloud platform replicates the weight parameters of the domain adaptation recognition model currently deployed on terminal devices as the baseline model for this optimization task. By aggregating expert knowledge and massive amounts of validated data, high-quality and comprehensive learning materials are provided for the model, and starting with a model in actual operation ensures the continuity and stability of the optimization.

[0092] Step S72: Incrementally train the baseline model based on the mixed dataset, and train the branch networks of the baseline model using sample data of the corresponding sample types to obtain the optimized model parameters.

[0093] As one implementation method, a mixed dataset is input into the model for forward and backward propagation calculations. During training, each training sample is automatically routed to its corresponding dedicated branch network for loss calculation and gradient backpropagation based on its type label. Specifically, fecal sample data is used to optimize the parameters of the fecal recognition branch network, while urine filter membrane sample data is specifically used to optimize the parameters of the urine recognition branch network. The mixed dataset is then input into the model for forward propagation, where sample images are fed into the baseline model and processed sequentially through each network layer to generate recognition results. The model's output is then compared with known correct answers, such as expert-corrected egg location and species, and the loss value is calculated using a loss function. After calculating the loss value through forward propagation, it is backpropagated back into the network. The error signal is then traced back to the shared backbone network through the corresponding dedicated branch network. During this process, the system calculates the direction and magnitude of adjustment required for each parameter in the backbone and branch networks, updates the model parameters accordingly, and obtains the optimized model parameters. By using closed-loop training with forward and backward propagation, combined with an automatic routing mechanism for sample types, precise optimization of the parameters of each dedicated branch network in the domain adaptation identification model was achieved.

[0094] In this embodiment, by constructing a hybrid dataset and using the domain-adaptive recognition model as the baseline model for domain-specific incremental training, precise optimization of the dedicated branch network is achieved, enabling the model to continuously learn from practical applications and constantly improve its recognition accuracy and adaptability to diverse samples.

[0095] For example, to help understand the technical concept or principle of the insect egg identification method of this embodiment, please refer to... Figure 7 , Figure 7 A simplified flowchart illustrating the method for identifying insect eggs in this application is shown below: First, samples are loaded onto the terminal device and quickly previewed. A lightweight classification network identifies the sample type, determining whether it is a fecal slide or a urine filter membrane. Based on the sample type, an adaptive scanning strategy and preprocessing algorithm are configured, such as enabling polarized light imaging and Fourier transform background suppression for urine filters, to complete high-resolution image acquisition and processing. The processed images are input into a domain-adaptive recognition model for inference. The model outputs the location information, classification result, and confidence score for each parasite egg. Based on the confidence threshold and classification clarity, an automatic judgment is made and a detection report is generated. Simultaneously, problematic images and their preliminary results are uploaded to the cloud platform. Cloud experts remotely review and correct problematic images, and incrementally train and optimize the model based on the corrected data and regular data. The optimized model parameters are then sent back to the terminal device in the form of differential updates, completing a complete collaborative closed loop of edge detection, cloud optimization, and terminal update. Through this complete collaborative closed loop between the terminal device and the cloud platform, automated high-precision detection of parasite eggs is achieved, with adaptive processing capabilities for different samples and continuous performance evolution capabilities.

[0096] Specifically, for example, in a village in an area endemic to schistosomiasis in Africa, testing personnel used a terminal device with a built-in battery to automatically scan urine filter membrane samples prepared according to the T / JPMA 014—2022 (Schistosoma haematobium egg detection urine filtration microscopy) standard. Utilizing edge computing capabilities, the terminal device completed image acquisition, background suppression, and AI recognition even without a network connection, generating a preliminary report. The terminal device uploaded the structured recognition results to a cloud platform via available mobile networks, such as 2G, 3G, or 4G, and could temporarily store the original images—that is, small datasets were uploaded to the cloud platform, while large datasets were temporarily stored. Experts on the cloud platform remotely reviewed and confirmed challenging samples, such as those with low confidence or suspected new species. The cloud platform aggregated high-quality data from multiple terminals worldwide, reviewed by experts, and regularly trained to generate better model versions. After connecting to a Wi-Fi network, the terminal device automatically downloads and updates the model, continuously evolving its recognition capabilities and completing a full collaborative optimization loop. This effectively solves the problem of limited resources for outdoor field detection while simultaneously achieving continuous improvement in model performance.

[0097] This application provides an insect egg identification system. The system employs a cloud-edge-device collaborative architecture and includes a terminal device and a cloud platform. The terminal device includes a microscopic imaging module, a motorized stage, an edge computing module, a battery module, a communication module, a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the insect egg identification method described above. The cloud platform includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the insect egg identification method described above.

[0098] Please refer to Figure 8 , Figure 8 This is a schematic diagram of the terminal device provided in this application. The terminal device integrates a microscopic imaging module, a motorized stage, an edge computing module with a built-in AI processing unit, a battery module, and a communication module supporting 4G, 5G, and Wi-Fi, forming a portable, off-grid-accessible field testing device. The microscopic imaging module includes a polarized light imaging component, which can automatically activate polarized light illumination mode for image acquisition of urine filtration membrane samples, thereby utilizing the characteristic that insect eggs have no birefringence while crystals have strong birefringence to distinguish crystals. Please refer to... Figure 9 , Figure 9 This is a schematic diagram of the overall architecture of the insect egg identification system provided in this application. The terminal device and the cloud platform interact and collaborate via a wireless communication network. The terminal device is responsible for automatic scanning of on-site samples, image preprocessing, and real-time identification of insect eggs based on the domain adaptation recognition model, and uploads the generated detection report, regular images, and difficult images, etc. The cloud platform receives the data, provides a remote expert review interface, and incrementally trains and optimizes the domain adaptation recognition model based on the data corrected by the expert review. Finally, the optimized model is distributed to the terminal device in a differential update manner, realizing distributed processing of detection tasks and continuous self-evolution of system capabilities.

[0099] The following is for reference. Figure 10 , Figure 10 This is a schematic diagram of the insect egg identification system in an embodiment of this application. Figure 10 The insect egg identification system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0100] like Figure 10As shown, the insect egg identification system may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the insect egg identification system. The processing device 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the insect egg identification system to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows an insect egg identification system with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0101] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0102] The insect egg identification system provided in this application, employing the insect egg identification method described in the above embodiments, can solve the technical problem of low accuracy in insect egg identification results. Compared with the prior art, the beneficial effects of the insect egg identification system provided in this application are the same as those of the insect egg identification system provided in the above embodiments, and other technical features of this insect egg identification system are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0103] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0105] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the insect egg identification method in the above embodiments.

[0106] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), or any suitable combination thereof.

[0107] The aforementioned computer-readable storage medium may be included in the insect egg identification system; or it may exist independently and not be assembled into the insect egg identification system.

[0108] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the insect egg identification system, the insect egg identification system performs background suppression on the acquired initial sample image to obtain a target sample image. The sample to be detected corresponding to the initial sample image includes fecal slide samples or urine filter membrane samples. The target sample image is input into a domain-adaptive identification model, which includes a backbone network and at least two branch networks. The backbone network is connected in parallel to at least two branch networks, with different sample types corresponding to different branch networks. The domain-adaptive identification model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg identification result. The insect egg identification result output by the domain-adaptive identification model is obtained, including location information, classification results, and confidence scores.

[0109] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0110] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0111] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0112] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described insect egg identification method, thereby solving the technical problem of low accuracy in insect egg identification results. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the insect egg identification method provided in the above embodiments, and will not be repeated here.

[0113] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the insect egg identification method described above.

[0114] The computer program product provided in this application can solve the technical problem of low accuracy in insect egg identification results. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the insect egg identification method provided in the above embodiments, and will not be repeated here.

[0115] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A method for identifying insect eggs, characterized in that, The insect egg identification method, applied to terminal devices, includes: Background suppression is performed on the acquired initial sample image to obtain the target sample image. The sample to be tested corresponding to the initial sample image includes a fecal slide sample or a urine filter membrane sample. The target sample image is input into a domain-adaptive recognition model, which includes a backbone network and at least two branch networks. The backbone network is connected in parallel to the at least two branch networks, and different sample types correspond to different branch networks. The domain-adaptive recognition model selects the corresponding branch network for inference based on the sample type of the target sample image to obtain the insect egg recognition result. Obtain the insect egg identification results output by the domain adaptation recognition model, wherein the insect egg identification results include location information, classification results and confidence level.

2. The method for identifying insect eggs as described in claim 1, characterized in that, The step of performing background suppression on the acquired initial sample image to obtain the target sample image includes: When the sample to be tested is the urine filter membrane sample, the initial sample image is converted from the spatial domain to the frequency domain to obtain a first frequency domain image, wherein the initial sample image is an image with a periodic grid texture; A band-stop filter is set according to the grid texture features of the filter membrane in the initial sample image, and the periodic grid background noise in the initial sample image is filtered out based on the band-stop filter to obtain the second frequency domain image; Perform an inverse Fourier transform on the second frequency domain image to convert it from the frequency domain to the spatial domain, thus obtaining the target sample image after background suppression.

3. The method for identifying insect eggs as described in claim 1, characterized in that, Before the step of performing background suppression on the acquired initial sample image to obtain the target sample image, the method further includes: The sample to be inspected is placed on the stage, and a preview scan is performed on the sample to be inspected and a preview image is acquired. The preview image is input into a preset classification network for recognition to obtain the sample type of the sample to be tested; The scanning strategy for the sample to be examined is determined based on the sample type, and the scanning strategy includes imaging mode and objective magnification. The sample to be inspected is scanned based on the scanning strategy to obtain the initial sample image.

4. The method for identifying insect eggs as described in claim 1, characterized in that, After the step of obtaining the insect egg identification result output by the domain adaptation recognition model, the method further includes: If the confidence level is greater than or equal to the confidence level threshold, and the ratio of the highest category probability to the second highest category probability in the classification result is higher than a preset ratio threshold, a detection report is generated based on the location information and the classification result; the confidence level is the probability of the presence of insect eggs, and the category probability is the probability corresponding to the classification result; The target sample image is labeled as a regular image.

5. The method for identifying insect eggs as described in claim 4, characterized in that, After obtaining the insect egg identification result output by the domain adaptation recognition model, the method further includes: When at least one preset condition is met, the target sample image is marked as a difficult image; Among them, meeting the preset conditions includes: The confidence level is less than the confidence threshold; When the overlapping area of ​​at least two detection boxes that mark the location of insect eggs is greater than a preset area threshold, and the highest classification probability of each detection box is higher than a preset classification threshold, the classification results corresponding to the highest classification probabilities of each detection box are inconsistent; the detection box is the bounding box marked by the domain adaptation recognition model in the target sample image for locating the location of insect eggs. The number of detection boxes output by the domain adaptation recognition model is 0, and the texture complexity of the target sample image is greater than the complexity threshold.

6. The method for identifying insect eggs as described in claim 5, characterized in that, The method further includes: The detection report, keyframe index, and regular image are uploaded to the cloud platform. The keyframe index is the image region number of the regular image where the insect egg is located. The problematic images and corresponding insect egg identification results are uploaded to the cloud platform; The system receives differential model parameters sent by the cloud platform and updates the parameters of the domain adaptation recognition model based on these differential model parameters.

7. A method for identifying insect eggs, characterized in that, The insect egg identification method, applied to a cloud platform, includes: The system receives a detection report, keyframe index, and corresponding regular image sent by the terminal device, as well as a difficult image and corresponding insect egg identification result. The keyframe index is the image region number of the regular image where the insect egg is located. The detection report, keyframe index, and corresponding regular image are used as regular samples; Obtain the manually verified and corrected insect egg identification results of the problematic image, and integrate the corrected insect egg identification results with the problematic image to obtain the problematic sample; Based on the regular samples and the difficult samples, the domain adaptation recognition model stored in the cloud platform is incrementally optimized to obtain the optimized model parameters. Determine the differential model parameters between the optimized model parameters and the domain adaptation recognition model parameters, and send the differential model parameters to the terminal device.

8. The method for identifying insect eggs as described in claim 7, characterized in that, The step of incrementally optimizing the domain adaptation recognition model stored in the cloud platform based on the regular samples and the difficult samples to obtain the optimized model parameters includes: The problematic samples and the regular samples are combined to form a hybrid dataset, with the domain adaptation recognition model as the baseline model. The baseline model is incrementally trained based on the hybrid dataset, and the branch networks of the baseline model are trained using sample data of the corresponding sample types to obtain the optimized model parameters.

9. A system for identifying insect eggs, characterized in that, The insect egg identification system adopts a cloud-edge-device collaborative architecture. The insect egg identification system includes a terminal device and a cloud platform. The terminal device includes: a microscopic imaging module, a motorized stage, an edge computing module, a battery module, a communication module, a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the insect egg identification method as described in any one of claims 1 to 6. The cloud platform includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the insect egg identification method as described in any one of claims 7 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the insect egg identification method as described in any one of claims 1 to 8.

Images