A pollen grain target detection system and method fusing multi-morphology priors

Abstract

The application provides a pollen grain target detection system and method fusing multiple morphological priors, and relates to the technical field of pollen detection. The application comprises the following steps: performing feature extraction on a pollen grain image to be detected to obtain an initial feature map, convolving the initial feature map using a pre-constructed morphological kernel library to obtain a morphological response feature map; performing adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map; processing a fusion feature obtained based on the morphological confidence map and a global context feature extracted from the initial feature map to obtain a spatial attention map; performing feature re-labeling on the initial feature map using the spatial attention map to obtain an enhanced feature map, and then performing target detection on the enhanced feature map to obtain pollen grain information output. The beneficial effects are as follows: effectively solving the problems of multiple background interferences and difficult morphological differentiation in pollen automatic detection, supporting high-time-efficiency pollen concentration monitoring, and overcoming the disadvantages of traditional manual microscopic examination processes, such as complexity, long cycle and dependence on professional personnel.

Description

Technical Field

[0001] This invention relates to the field of pollen detection technology, and in particular to a pollen grain target detection system and method that integrates multiple morphological priors. Background Technology

[0002] Allergenic pollen is a major trigger for seasonal allergic rhinitis and asthma. Accurate detection of pollen types and concentrations in the air is crucial for meteorological departments to issue pollen warnings and guide susceptible populations in effective protection. Currently used pollen monitoring methods (such as the Hirst method) mainly rely on manual morphological identification and counting of samples (such as glass slides) under a microscope. This method is complex, time-consuming, and requires a high level of professional morphological knowledge from the personnel, making it difficult to achieve efficient, rapid, and automated monitoring.

[0003] Automated object detection technologies, such as convolutional neural networks (CNNs), have offered new possibilities for automatic pollen identification. However, directly applying general object detection models to pollen grain identification still faces the following specific challenges:

[0004] 1. Complex background interference: In addition to the target pollen, the air sample contains a large number of non-target particles, such as fungal spores, dust, fibers and other biological debris. They are very similar to some pollen in shape and size, which can easily cause interference.

[0005] 2. Significant similarity between species and significant differences within species: Pollen from different families and genera may be highly similar in morphology (e.g., Populus and Salix), while the same species of pollen may present different appearances or postures due to different collection conditions, dehydration state, or microscope focal plane.

[0006] 3. Lack of morphological prior constraints: Existing deep learning models usually do not explicitly incorporate the inherent morphological features of pollen (such as spherical, elliptical, polygonal or air sac-like structures), which makes the model susceptible to interference from similar textures in the background, resulting in false detections and localization biases, leading to a high false alarm rate and inaccurate boundary localization.

[0007] Therefore, how to effectively combine prior knowledge about pollen morphology in aerobic biology with the powerful feature extraction capabilities of deep learning to construct an automated detection system that can accurately distinguish between pollen and non-pollen particles and achieve fine-grained pollen classification is a key issue that urgently needs to be addressed in this field. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a pollen grain target detection system that integrates multiple morphological priors, comprising: a morphological feature extraction module, used to extract features from an image of pollen grains to be detected to obtain an initial feature map, and to convolve the initial feature map using a pre-built morphological convolution kernel library to obtain a morphological response feature map; a guided attention module, connected to the morphological feature extraction module, used to perform adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, and then to process the fused features obtained by combining the morphological confidence map with global context features extracted from the initial feature map to obtain a spatial attention map; and a detection module, connected to the morphological extraction module, used to recalibrate the initial feature map using the spatial attention map to obtain an enhanced feature map, and then to perform target detection on the enhanced feature map to obtain pollen grain information output.

[0009] Preferably, a variety of basic morphological nuclei corresponding to pollen grains are preset, and then the morphology of each basic morphological nucleus at multiple different scales and multiple different directions is generated by mathematical equations to obtain multiple morphological convolution kernels, so as to form the morphological convolution kernel library.

[0010] Preferably, the morphological feature extraction module includes: an initial extraction unit, used to acquire a microscopic image or video frame containing pollen grains as the pollen grain image, and then input the pollen grain image into a feature extraction network to extract the initial feature map; and a morphological extraction unit, connected to the initial extraction unit, used to convolve the initial feature map using each morphological convolution kernel in the morphological convolution kernel library to generate a set of morphological response feature maps.

[0011] Preferably, the guided attention module includes: a confidence map generation unit, used to input the morphological response feature map into a convolutional network with adaptive weights for feature fusion to obtain the morphological confidence map; a global feature extraction unit, used to compute the global average pooling and global max pooling of the initial feature map in parallel as the global context features; and a feature fusion unit, connected to the confidence map generation unit and the global feature extraction unit respectively, used to concatenate the global context features and the morphological confidence map, and then input them into a spatial fusion network with an activation function applied for feature fusion to obtain the spatial attention map.

[0012] Preferably, the detection module includes: a recalibration unit, used to multiply the spatial attention map as a mask or weight with the initial feature map element-wise, and combine it with residual connections to obtain the enhanced feature map; and a target detection unit, connected to the recalibration unit, used to input the enhanced feature map into the detection head network to decode the enhanced feature map, obtain the number of pollen grains contained in the pollen grain image, and obtain the bounding box, confidence score and pollen category of each pollen grain as the pollen grain information output.

[0013] Preferably, it also includes an interaction module, which connects the detection module, the morphological feature extraction module and the guided attention module, for allowing operators to view the morphological response feature map, the morphological confidence map, the spatial attention map and the pollen grain information, and for allowing operators to simultaneously compare the pollen grain information with the detection results of multiple pre-configured comparison models.

[0014] This invention also provides a pollen grain target detection method that integrates multiple morphological priors, applied to the aforementioned pollen grain target detection system, comprising: step S1, the pollen grain target detection system extracts features from the pollen grain image to be detected to obtain an initial feature map, and convolves the initial feature map using a pre-built morphological convolution kernel library to obtain a morphological response feature map; step S2, the pollen grain target detection system performs adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, and then performs fusion feature processing based on the morphological confidence map and global context features extracted from the initial feature map to obtain a spatial attention map; step S3, the pollen grain target detection system recalibrates the initial feature map using the spatial attention map to obtain an enhanced feature map, and then performs target detection on the enhanced feature map to obtain pollen grain information output.

[0015] Preferably, the pollen grain target detection system pre-sets multiple basic morphological kernels corresponding to pollen grains, and then generates multiple morphological convolution kernels by mathematical equations at multiple different scales and multiple different directions to form the morphological convolution kernel library.

[0016] Preferably, step S1 includes: step S1, the pollen grain target detection system acquires a microscopic image or video frame containing pollen grains as the pollen grain image, and then inputs the pollen grain image into a feature extraction network to extract the initial feature map; step S12, the pollen grain target detection system uses each morphological convolution kernel in the morphological convolution kernel library to convolve the initial feature map to generate a set of morphological response feature maps.

[0017] Preferably, step S2 includes: step S21, the pollen grain target detection system inputs the morphological response feature map into a convolutional network with adaptive weights for feature fusion to obtain the morphological confidence map; step S22, the pollen grain target detection system calculates the global average pooling and global max pooling of the initial feature map in parallel as the global context features; step S23, the pollen grain target detection system concatenates the global context features and the morphological confidence map, and then inputs them into a spatial fusion network with an applied activation function for feature fusion to obtain the spatial attention map.

[0018] The above technical solution has the following advantages or beneficial effects: The system of the present invention deeply integrates prior knowledge of pollen morphology into the target detection network, constructs a morphology-guided attention mechanism, effectively solves the core problems of background interference and difficulty in distinguishing morphological differences in automatic pollen detection, and provides reliable technical support for high-timeliness pollen concentration monitoring. It also overcomes the shortcomings of traditional manual microscopic examination, such as cumbersome process, long cycle, and high dependence on professional personnel. Attached Figure Description

[0019] Figure 1 A schematic diagram of a pollen grain target detection system that integrates multiple morphological priors is shown in a preferred embodiment of the present invention.

[0020] Figure 2 A flowchart illustrating a pollen grain target detection method that integrates multiple morphological priors is shown in a preferred embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of a sub-process of step S1 in a preferred embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the sub-process of step S2 in a preferred embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0024] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a pollen grain target detection system integrating multiple morphological priors is provided, such as... Figure 1As shown, it includes: a morphological feature extraction module 1, used to extract features from the pollen grain image to be detected to obtain an initial feature map, and to convolve the initial feature map using a pre-built morphological convolution kernel library to obtain a morphological response feature map; a guided attention module 2, connected to the morphological feature extraction module 1, used to perform adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, and then to perform fusion feature processing based on the morphological confidence map and the global context features extracted from the initial feature map to obtain a spatial attention map; and a detection module 3, connected to the morphological extraction module 2, used to recalibrate the initial feature map with the spatial attention map to obtain an enhanced feature map, and then to perform target detection on the enhanced feature map to obtain pollen grain information output.

[0025] Specifically, in this embodiment, the morphological feature extraction module 1 performs convolution operations on the initial feature map using a pre-built library of morphological convolution kernels (such as circular, elliptical, rectangular, and rhomboid kernels) to generate a set of morphological response feature maps. These morphological response feature maps characterize the response strength of the initial feature map to different morphological priors.

[0026] This explicitly incorporates prior knowledge of typical pollen grain geometry (e.g., spherical, oval, polygonal), enabling the model to directly capture the inherent shape features of pollen. It addresses the limitation of existing deep learning models being easily misled by background pseudo-features (e.g., dust, spores). Through the enhancement of morphological priors, the model can more accurately identify pollen regions, reduce false alarm rates, and improve the robustness of feature extraction.

[0027] Furthermore, the attention module 2 performs adaptive weighted feature fusion based on the morphological response feature map to generate a morphological confidence map. This morphological confidence map highlights regions that match the pre-defined pollen morphologies in the morphological convolutional kernel library. It then combines the global contextual features of the initial feature map (such as global average pooling and max pooling features) to obtain a spatial attention map. This achieves proactive focusing on pollen morphology regions while suppressing feature responses from non-pollen interferences. The attention mechanism strengthens pollen feature regions, reducing the risk of misclassification due to inter-class similarity. For example, for pollen from different genera and families with similar morphologies (such as Populus and Salix), this module provides fine-grained shape constraints, improving classification accuracy.

[0028] Detection module 3 uses a spatial attention map to recalibrate the initial feature map, resulting in an enhanced feature map. This enhanced feature map is then output through the detection head network, providing pollen grain information (such as bounding boxes, confidence scores, and categories). This enables the model to more accurately locate pollen grain boundaries and improves detection confidence. This module amplifies pollen region features through feature recalibration and suppresses background interference, thereby improving detection accuracy and localization reliability. Furthermore, this module achieves automated detection, solving the problems of low efficiency and long processing times associated with traditional manual microscopic examination.

[0029] In summary, through the collaborative work of the aforementioned modules and guided by morphological priors, this system can effectively distinguish pollen grains from morphologically similar non-pollen debris (such as fungal spores and dust), significantly reducing the false alarm rate. For pollen from different species with similar morphologies, the system utilizes a morphological attention mechanism to provide fine-grained shape constraints in addition to texture and color, which helps improve inter-class differentiation and reduce misclassification.

[0030] Overall, the system deeply integrates prior knowledge of pollen morphology into the target detection network and constructs a morphology-guided attention mechanism, effectively solving the core problems of background interference and difficulty in distinguishing morphological differences in automatic pollen detection, providing reliable technical support for high-timeliness pollen concentration monitoring. It also overcomes the shortcomings of traditional manual microscopic examination, such as cumbersome procedures, long cycles, and high dependence on professional personnel.

[0031] In a preferred embodiment of the present invention, a variety of basic morphological nuclei corresponding to pollen grains are preset, and then the morphology of each basic morphological nucleus at multiple different scales and multiple different directions is generated by mathematical equations to obtain multiple morphological convolution kernels, so as to form a morphological convolution kernel library.

[0032] Specifically, the multimorphic convolutional kernel library in this embodiment is designed based on the morphological features of target pollen (such as Artemisia, Platanus, and Pine). Its core function is to accurately match typical pollen shapes such as spherical, oval, and polygonal, and it covers multi-scale and multi-directional requirements. Specifically, the convolutional kernels in the library are generated through mathematical equations corresponding to preset morphological features of the pollen—preset morphological features include circular, elliptical, rectangular, and rhomboid shapes (at least one), and the corresponding mathematical equations cover elliptical equations, Manhattan distance, Chebyshev distance, etc. To address the morphological diversity of pollen grains, the kernel generator is further configured with four basic morphological kernels, each with the following applicable scenarios:

[0033] 1. Elliptical nucleus: suitable for oval pollen from Artemisia, Chenopodiaceae, and other similar species;

[0034] 2. Circular core: Suitable for spherical pollen from species such as Platanus and Salicaceae;

[0035] 3. Rectangular core: adapted to partially folded or elongated pollen grains viewed from a specific angle;

[0036] 4. Diamond: Compatible with some pine family (such as fir) or polygonal cypress family pollen.

[0037] Each morphology generates kernels at 4 scales and in 8 directions (0° to 150°), for a total of 128 morphological convolution kernels. .

[0038] In a preferred embodiment of the present invention, such as Figure 1 As shown, the morphological feature extraction module 1 includes: an initial extraction unit 11, used to acquire a microscopic image or video frame containing pollen grains as a pollen grain image, and then input the pollen grain image into the feature extraction network to extract an initial feature map; and a morphological extraction unit 12, connected to the initial extraction unit 11, used to convolve the initial feature map using each morphological convolution kernel in the morphological convolution kernel library to generate a set of morphological response feature maps.

[0039] Specifically, in this embodiment, the system inputs a pollen detection slide image (640x640 pixels). This image can be a microscopic image directly captured during pollen observation or a video frame extracted from a microscope video stream. An initial feature map is extracted by a feature extraction network (preferably a deep convolutional neural network, such as a YOLOv8 backbone network, is used in this embodiment). .

[0040] use right Convolution is performed to obtain a 128-channel morphological response map. This set of feature maps characterizes the response strength of the initial feature maps to different morphological priors.

[0041] In a preferred embodiment of the present invention, such as Figure 1 As shown, the guided attention module 2 includes: a confidence map generation unit 21, used to input the morphological response feature map into a convolutional network with adaptive weights for feature fusion to obtain the morphological confidence map; a global feature extraction unit 22, used to compute the global average pooling and global max pooling of the initial feature map in parallel as global context features; and a feature fusion unit 23, which connects the confidence map generation unit 21 and the global feature extraction unit 22 respectively, used to concatenate the global context features and the morphological confidence map, and then input them into a spatial fusion network with an activation function applied for feature fusion to obtain a spatial attention map.

[0042] Specifically, in this embodiment, the morphological response feature map is used. The input is fed into a feature fusion network, which uses adaptive weighting (a 1×1 convolutional layer with adaptive weights (learnable parameters) is used as the feature fusion network, and the fusion weights of different morphological features in the morphological response feature map are determined through an adaptive weight learning mechanism) to convert the morphological response feature map into a single feature. Merged into a single-channel morphological confidence map The image highlights the regions that match the preset pollen morphology in the morphological convolution kernel library, which is used to characterize the degree of matching between the spatial location in the initial feature map and the preset morphological features.

[0043] Furthermore, parallel computation Global average pooling and global max pooling .Will , and After concatenation, the data is fed into a spatial fusion network (7×7 convolutional layers with a sigmoid activation function) to generate the final spatial attention map. .

[0044] In a preferred embodiment of the present invention, such as Figure 1 As shown, the detection module 3 includes: a recalibration unit 31, which is used to multiply the spatial attention map as a mask or weight with the initial feature map element-wise, and combine it with residual connections to obtain an enhanced feature map; and a target detection unit 32, connected to the recalibration unit 31, which is used to input the enhanced feature map into the detection head network to decode the enhanced feature map, obtain the number of pollen grains contained in the pollen grain image, and obtain the bounding box, confidence score and pollen category of each pollen grain as pollen grain information output.

[0045] Specifically, in this embodiment, feature recalibration is performed by weighting the initial feature map using a spatial attention map to obtain an enhanced feature map. The specific process is as follows: the spatial attention map... As a mask or weight, it is used in conjunction with the initial feature map. Element-wise multiplication and residual connections are performed to obtain the enhanced feature map. :

[0046] = + ;

[0047] The pollen grains are then fed into the YOLOv8 neck network (PANet) and the detector head network for decoding and output. The output includes pollen grain information, specifically the bounding box and location of the pollen grains, confidence level, and specific pollen category (e.g., Artemisia pollen, Platanus pollen, Birch pollen, etc.), and the pollen grain count is completed in real time.

[0048] In a preferred embodiment of the present invention, an interaction module is further included, which connects the detection module, the morphological feature extraction module, and the guided attention module. This module is used to allow operators to view the morphological response feature map, the morphological confidence map, the spatial attention map, and the pollen grain information. It is also used to allow operators to simultaneously compare the pollen grain information with the detection results of multiple pre-configured comparison models.

[0049] Specifically, in this embodiment, the interaction module can implement multiple functions, including:

[0050] 1. Information viewing function

[0051] Operators can view morphological response feature maps and morphological confidence maps. Heatmaps and spatial attention maps The heatmap visually presents the regions in the image that match the preset pollen morphology and the intensity of their feature responses; it also allows users to view a comparison of feature maps before and after the application of the attention mechanism, providing a clear understanding of the feature enhancement effect.

[0052] It provides a real-time counting panel for 14 common allergenic pollen types (supporting Chinese and English bilingual switching), simultaneously displaying the pollen grain bounding box, confidence level, and specific category (such as Artemisia pollen, Platanus pollen, etc.) of the detection output, and supports operators to manually edit the detection quantity for manual correction.

[0053] 2. Multi-model detection result synchronous comparison function

[0054] It supports loading two or more different pre-configured comparison models (such as the baseline YOLO model and the model of this invention) to process the same video source, keeping them synchronized through the syncVideos function, and allowing for intuitive comparison of detection differences. Through synchronized playback, pause, progress bar dragging, and other operations, the differences in detection results between different models can be intuitively compared.

[0055] It can generate and view comparison charts of detection boxes from multiple models, plotting the actual values ​​and the predicted values ​​of different models on the same chart to help operators analyze differences in model detection; it also supports comparing morphological response heatmaps and attention comparison charts corresponding to different models to gain a deeper understanding of the differences in the model's decision-making basis.

[0056] 3. Parameter configuration and interactive adjustment functions

[0057] It supports operators to adjust the core parameters of models such as the confidence threshold, IOU threshold, and maximum number of detections in real time, including the detection head network, feature extraction network, and spatial fusion network. The detection results are fed back immediately after the adjustment to optimize the detection accuracy.

[0058] Operators can choose input size, batch size, and data augmentation strategy, and can dynamically switch between loading different pre-trained model weights to flexibly adapt to different detection scenario requirements.

[0059] 4. Results export function

[0060] It supports manual single-frame screenshot and automatic batch screenshot functions. It uses the HTML5 Canvas API to capture image data from video frames. Operators can export the viewed morphological feature maps, attention maps, detection result comparison maps, etc. in high-resolution PNG format of 300 DPI for easy subsequent analysis and archiving.

[0061] This invention also provides a pollen grain target detection method that integrates multiple morphological priors, applied to the aforementioned pollen grain target detection system, such as... Figure 2 As shown, the process includes: Step S1, the pollen grain target detection system extracts features from the pollen grain image to be detected to obtain an initial feature map, and convolves the initial feature map using a pre-built morphological convolution kernel library to obtain a morphological response feature map; Step S2, the pollen grain target detection system performs adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, and then performs fusion feature processing based on the morphological confidence map and the global context features extracted from the initial feature map to obtain a spatial attention map; Step S3, the pollen grain target detection system recalibrates the initial feature map with the spatial attention map to obtain an enhanced feature map, and then performs target detection on the enhanced feature map to obtain pollen grain information output.

[0062] In a preferred embodiment of the present invention, the pollen grain target detection system pre-sets a variety of basic morphological kernels corresponding to pollen grains, and then generates the morphology of each basic morphological kernel at multiple different scales and multiple different directions through mathematical equations to obtain multiple morphological convolution kernels, so as to form a morphological convolution kernel library.

[0063] In a preferred embodiment of the present invention, such as Figure 3 As shown, step S1 includes: step S11, the pollen grain target detection system acquires a microscopic image or video frame containing pollen grains as a pollen grain image, and then inputs the pollen grain image into the feature extraction network to extract an initial feature map; step S12, the pollen grain target detection system uses each morphological convolution kernel in the morphological convolution kernel library to convolve the initial feature map to generate a set of morphological response feature maps.

[0064] In a preferred embodiment of the present invention, such as Figure 4 As shown, step S2 includes: step S21, the pollen grain target detection system inputs the morphological response feature map into a convolutional network with adaptive weights for feature fusion to obtain the morphological confidence map; step S22, the pollen grain target detection system calculates the global average pooling and global max pooling of the initial feature map in parallel as global context features; step S23, the pollen grain target detection system concatenates the global context features and the morphological confidence map, and then inputs them into a spatial fusion network with an applied activation function for feature fusion to obtain a spatial attention map.

[0065] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A pollen grain object detection system fusing multi-morphology priors, characterized in that, The method comprises the following steps: a morphological feature extraction module is used to extract features from a pollen grain image to be detected to obtain an initial feature map, and a pre-constructed morphological convolution kernel library is used to convolve the initial feature map to obtain a morphological response feature map; a guided attention module is connected to the morphological feature extraction module, and is used to perform adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, then perform feature fusion based on the morphological confidence map and global context features extracted from the initial feature map, and then process the fused features to obtain a spatial attention map; a detection module is connected to the morphological extraction module, and is used to perform feature re-labeling on the initial feature map by using the spatial attention map to obtain an enhanced feature map, and then perform target detection on the enhanced feature map to obtain pollen grain information output.

2. The pollen grain target detection system of claim 1, wherein, A plurality of basic morphological kernels corresponding to a plurality of pollen grains are preset, and then a plurality of morphological convolution kernels are generated by mathematical equations to form the morphological convolution kernel library.

3. The pollen grain target detection system of claim 1, wherein, The morphological feature extraction module comprises: an initial extraction unit is used to obtain a microscopic image or a video frame containing pollen grains as the pollen grain image, and then input the pollen grain image into a feature extraction network to extract the initial feature map; a morphological extraction unit is connected to the initial extraction unit, and is used to convolve the initial feature map using each morphological convolution kernel in the morphological convolution kernel library to generate a group of morphological response feature maps.

4. The pollen grain target detection system of claim 1, wherein, The guided attention module comprises: a confidence map generation unit is used to input the morphological response feature map into a convolution network with adaptive weights to perform feature fusion to obtain the morphological confidence map; a global feature extraction unit is used to calculate the global average pooling and global maximum pooling of the initial feature map in parallel as the global context features; a feature fusion unit is connected to the confidence map generation unit and the global feature extraction unit respectively, and is used to concatenate the global context features and the morphological confidence map, and then input into a spatial fusion network with an activation function to perform feature fusion to obtain the spatial attention map.

5. The pollen grain target detection system of claim 1, wherein, The detection module comprises: a re-labeling unit is used to multiply the spatial attention map as a mask or weight with the initial feature map element by element, and combine a residual connection to obtain the enhanced feature map; a target detection unit is connected to the re-labeling unit, and is used to input the enhanced feature map into a detection head network to decode the enhanced feature map, obtain the number of pollen grains contained in the pollen grain image, and obtain the bounding box, confidence and pollen category of each pollen grain as the pollen grain information output.

6. The pollen grain target detection system of claim 1, wherein, Further comprising an interaction module connected to the detection module, the morphological feature extraction module and the attention guiding module, for the operator to view the morphological response feature map, the morphological confidence map, the spatial attention map and the pollen grain information, and also for the operator to synchronously compare the pollen grain information with the detection results of multiple preconfigured comparison models.

7. A pollen grain object detection method fusing multi-morphology priors, characterized in that, The pollen grain target detection system is applied to the pollen grain target detection system in any one of claims 1-6, comprising: Step S1, the pollen grain target detection system extracts features from the pollen grain image to be detected to obtain an initial feature map, and uses a pre-constructed morphological convolution kernel library to convolve the initial feature map to obtain a morphological response feature map; Step S2, the pollen grain target detection system performs adaptive weighted feature fusion based on the morphological response feature map to obtain a morphological confidence map, and then performs feature fusion based on the morphological confidence map and global context features extracted from the initial feature map, and then processes the fusion features to obtain a spatial attention map; Step S3, the pollen grain target detection system performs feature re-calibration on the initial feature map using the spatial attention map to obtain an enhanced feature map, and then performs target detection on the enhanced feature map to obtain pollen grain information output.

8. The pollen grain target detection method according to claim 7, characterized by, The pollen grain target detection system predefines a plurality of basic morphological kernels corresponding to a plurality of pollen grains, and then generates a plurality of morphological convolution kernels by mathematical equations, so as to form the morphological convolution kernel library.

9. The pollen grain target detection method according to claim 7, characterized by, The step S1 comprises: Step S11, the pollen grain target detection system acquires a microscopic image or a video frame containing pollen grains as the pollen grain image, and then inputs the pollen grain image into a feature extraction network to extract the initial feature map; Step S12, the pollen grain target detection system uses each morphological convolution kernel in the morphological convolution kernel library to convolve the initial feature map to generate a group of morphological response feature maps.

10. The pollen grain target detection method according to claim 7, characterized by, The step S2 comprises: Step S21, the pollen grain target detection system inputs the morphological response feature map into a convolution network with adaptive weights to perform feature fusion and obtain the morphological confidence map; Step S22, the pollen grain target detection system calculates the global average pooling and the global maximum pooling of the initial feature map in parallel as the global context features; Step S23, the pollen grain target detection system splices the global context features and the morphological confidence map, and then inputs them into a spatial fusion network with an activation function to perform feature fusion and obtain the spatial attention map.

Images