A blood cell automatic detection method based on semi-supervised learning

Abstract

This invention relates to an automatic blood cell detection method based on semi-supervised learning, belonging to the field of intelligent medical image analysis. The invention is achieved through the following technical solution: First, dual parallel detection networks perform detection predictions separately, generating their respective pseudo-label results; then, based on the dual criteria of prediction confidence and localization consistency, high-quality detection instances are selected and stored in a dynamic instance pool, which continuously maintains the pseudo-labels of instances acquired in the most recent training rounds; next, a category-scale balance factor fusing category frequency and target scale is introduced to allocate sampling weights to high-quality instances in the dynamic instance pool, and instances are randomly selected for instance-pasting data augmentation; furthermore, during instance pasting, a background-aware pasting strategy and a location-random sampling mechanism are adopted to accurately paste the sampled instances into the background area of ​​the image, thereby guiding the detection network to focus on learning the discriminative features of a few categories and small-scale targets, significantly improving its detection performance.

Description

Technical Field

[0001] This invention relates to a medical microscopic image target detection technology, particularly a semi-supervised target detection method based on deep neural networks in blood cell microscopic images, belonging to the field of intelligent medical image analysis. Background Technology

[0002] With the continuous development of microscopic imaging equipment, blood cell detection and analysis based on microscopic images are playing an increasingly important role in the auxiliary diagnosis of blood diseases. By automatically detecting and locating cells such as red blood cells, white blood cells, and platelets in blood microscopic images, quantitative analysis of blood components can be achieved, providing important evidence for the diagnosis and treatment of various hematological diseases such as anemia, infectious diseases, and leukemia. However, traditional methods relying on manual microscopic observation and counting are not only inefficient but also highly dependent on physician experience, making it difficult to meet the needs of large-scale clinical applications. In recent years, deep learning-based blood cell detection models can automatically learn discriminative features from a large number of labeled microscopic images, outperforming traditional methods in both detection accuracy and efficiency.

[0003] Currently, most high-performance detection models rely on large amounts of high-quality labeled data for fully supervised training. However, accurate labeling of blood cell microscopic images also requires personnel with professional medical backgrounds to label each cell instance individually, resulting in extremely high time and manpower costs, limiting the widespread application of these methods in real-world clinical settings. To reduce reliance on manually labeled data, semi-supervised object detection methods have gradually gained attention. These methods alleviate the high labeling costs to some extent by jointly utilizing a small amount of labeled data with a large amount of unlabeled data for model training. However, in the specific application scenario of blood cell detection, existing semi-supervised object detection methods still face significant challenges. On the one hand, blood cell microscopic images typically contain a wide variety of cell types in a single image, such as red blood cells being far more numerous than white blood cells and platelets, leading to a severe imbalance in the distribution of these categories. On the other hand, different types of blood cells differ significantly in spatial scale; platelets are extremely small and scarce, while white blood cells are larger, and this scale imbalance further exacerbates the detection difficulty.

[0004] An examination of current object detection methods reveals that while they achieve high performance in blood cell detection, they rely on training with large amounts of labeled images, resulting in poor performance in scenarios where training labels are difficult to obtain in large quantities. Existing semi-supervised object detection methods typically rely on consistency constraints and pseudo-label learning to handle large amounts of unlabeled images, but they usually assume a relatively balanced class distribution and target scale. When directly applied to blood cell microscopic images, these methods often struggle to effectively balance minority classes and small-scale targets, leading to decreased pseudo-label quality, unstable model training, and consequently affecting detection performance. Furthermore, existing methods often fail to effectively model and utilize unlabeled data at both the class and scale levels, making it difficult to fully extract valuable information about minority classes and small targets from unlabeled data. Therefore, new methods are needed to address the class and scale imbalance problems in semi-supervised blood cell detection tasks, improving the detection accuracy of blood cells while reducing manual annotation costs. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of existing fully supervised blood cell detection methods, which heavily rely on extensive manual annotation and are costly, and the shortcomings of existing semi-supervised target detection methods in detecting a few categories and small-scale targets in microscopic images of blood cells. This invention proposes a semi-supervised target detection method that balances category and scale for detecting multiple cell categories in blood cell microscopic images. This method achieves collaborative training by constructing a dual-branch parallel detection network architecture enhanced with high-confidence instances, and introduces a category and scale balancing mechanism on unlabeled data, effectively improving the detection capability of a few categories and small-scale targets.

[0006] This invention proposes a novel semi-supervised learning-based target detection method to address the common problems of high annotation costs, uneven class distribution, and significant differences in target scale in cell microscopy images. This is achieved through the following technical solution: For unlabeled blood cell microscopy images, a dual-branch parallel detection network first performs detection prediction to generate preliminary pseudo-labels. Then, based on both prediction confidence and localization consistency criteria, high-quality detection instances are selected and stored in a dynamic instance pool, which continuously maintains high-quality samples acquired in recent training rounds. Next, a class-scale balance factor fusing class frequency and target scale is introduced to allocate sampling weights to high-quality instances in the dynamic instance pool, and instances are randomly selected for instance-pasting data augmentation. Furthermore, during instance pasting, a background-aware pasting strategy and a random location sampling mechanism are employed to accurately paste sampled instances into the background region of the image, effectively avoiding damage to the structural integrity and semantic context consistency of the original image. This guides the detection network to focus on learning the discriminative features of low-frequency and small-scale targets, significantly improving its detection performance.

[0007] The technical solution of this invention is an automated blood cell detection method based on semi-supervised learning, the method comprising:

[0008] Step 1: Based on dual-view Figure 1 Building a consistent, high-quality instance pool;

[0009] In a training batch, containing Image of blood cells with annotations and its annotation , Zhang's unlabeled blood cell image For a single labeled image Its labeling information Include One foreground cell instance; One-hot encoding representing category labels, Represents the bounding box coordinates. It is a collection of cell categories;

[0010] Use two detectors with identical structures but different initialization parameters. and As two different views, they learn through a pseudo-label cross-supervision mechanism on unlabeled images;

[0011] Adopting a dual-view-based approach Figure 1 A consistent pseudo-tag quality scoring and dynamic instance pool construction method is used to evaluate the quality of pseudo-tags and construct an instance pool using high-quality instance-level pseudo-tags.

[0012] Step 2: Sampling for class and scale balance in the instance pool;

[0013] Instance samples are randomly sampled from the instance pool and pasted onto the original image to generate two different versions of enhanced images, which are then used as... and Input;

[0014] For the generation detector Taking the corresponding enhanced image as an example, firstly, for each batch of images, the category sampling probability is set according to the category frequency. as follows:

[0015] Formula 4;

[0016] Formula 5;

[0017] in This represents a weighted score that combines category frequency with size. Indicates detector Category in the predicted reliable pseudo-labels The reciprocal of the frequency, For set The average area of ​​the middle bounding box; Used to control the degree of contribution of scale information. It is a constant less than the set threshold to avoid division by zero error; the category sampling probability defined based on Formula 4 can improve the sampling probability of rare categories and small-scale categories, thereby alleviating the imbalance problem;

[0018] Secondly, based on probability After determining the category to be sampled, samples are adaptively and randomly selected from instances of that category for subsequent pasting operations; to ensure that the number of cells in the pasted image is within a reasonable range, a maximum allowed number of instances is introduced. This parameter represents the maximum number of cells contained in any image of the labeled subset;

[0019] For those with false labels Single unlabeled image The total number of samples that need to be sampled for:

[0020] Formula 6;

[0021] This represents the predefined maximum number of instances for each category. From the dynamic pool respectively Medium sampling Each image has one sample; the final sampled instance set for each image is... ,in and These represent instance image patches and their corresponding instances, respectively.

[0022] Step 3: Based on the instance set obtained from the above steps These instances are then pasted into the original image for data augmentation;

[0023] Step 4: Cross-supervised learning based on instance-based augmented data:

[0024] For two detectors and pseudo-tags and Perform steps 2 and 3 independently to obtain instance augmentation data pairs. and ; to detector Pseudo-labels are used for monitoring detectors At the same time, the detector Pseudo-labels are used for monitoring detectors Calculate the cross-pseudo-label supervision loss between the two. and :

[0025] Formula 10;

[0026] Formula 11;

[0027] in, This indicates the number of unlabeled images in the training set. , Indicates detector-based and The predicted pseudo-labels are then processed in steps 2 and 3 to obtain the instance-enhanced image. , This represents the pseudo-tag corresponding to the instance enhancement. For classifying losses, The bounding box regression loss; in addition, the detector and Calculate the supervision loss on the labeled data respectively. :

[0028] Formula 12;

[0029] The total loss during training is defined as:

[0030] Formula 13;

[0031] in, The weights are used to adjust the proportion of the unsupervised loss component in the total loss. Based on the above loss function definition, the detector parameters are updated through backpropagation to obtain the trained detection model. and ;

[0032] Step 5: During the testing phase, a blood cell test image is used. Using a trained detector Prediction is performed to obtain blood cell test results. Furthermore, non-maximum suppression is used to merge detection instances with consistent categories and high overlap (cross-union ratio greater than a threshold) into a single detection instance.

[0033] Furthermore, step 1 employs a dual-view-based approach. Figure 1 The specific method for evaluating the quality of pseudo-tags and utilizing high-quality instance-level pseudo-tags is as follows: (This is based on a consistent pseudo-tag quality scoring and dynamic instance pool construction method.)

[0034] First, the unlabeled image Input detectors respectively and Generate the corresponding initial pseudo-tags. and ; Medium category The set of cells is denoted as Its size is recorded as ; Medium category The set of cells is denoted as Its size is recorded as ; for categories The intersection-union ratio (CUI) is calculated for any pair of bounding boxes of two targets in these two subsets, resulting in an CUI matrix. ,in represent The A detection box and The The IoU value of each detection box; based on Filter out noisy instances by selecting bidirectional matching instance pairs with consistent positioning.

[0035] Formula 1;

[0036] in , ; It is the set of instance pairs that match the predictions of the two networks after filtering; and They represent The and the One detection box, , They represent The The and the first One detection box;

[0037] For each bidirectional matching pair Define a quality score that integrates classification confidence and localization consistency:

[0038] Formula 2;

[0039] in Represents the predicted category probability distribution; due to and To represent the same instance, assign it the same quality score. ;

[0040] Secondly, a quality scoring threshold is set based on this. Only retain those with a quality score higher than [the previous one]. Instances, as a collection of high-quality instances. :

[0041] Formula 3;

[0042] Next, using Build an instance pool from high-quality instances. The length of this instance pool is fixed. Considering that minority class instances are extremely scarce within each image, to alleviate the class imbalance problem in the instance pool, from each class... One instance is randomly selected from the pool and added to the dynamic instance pool. At the end, A first-in-first-out (FIFO) dynamic update strategy is adopted to maintain only the latest samples to ensure data timeliness.

[0043] Furthermore, in step 2 .

[0044] Furthermore, the specific method of step 3 is to use an enhanced background-aware pasting method, which includes:

[0045] First, for unlabeled images Gamma correction is performed to enhance the contrast between the foreground and background of the image. An adaptive threshold is calculated using the large-law method to generate a binary background mask, thereby obtaining the pixel set of the background region. For each in the set Pixels Calculate the minimum Euclidean distance to the foreground-background boundary. A selection probability is defined based on this distance, such that pixels farther from the boundary have a higher selection probability:

[0046] Formula 7;

[0047] Secondly, for instance collections Each instance in Based on probability distribution Randomly sample a pixel Using this instance as the center, paste it to the corresponding position in the image; after the pasting operation is complete, based on the pixel... Position recalibration instance The bounding box coordinates, and the calibrated set of instances are denoted as ;

[0048] At the same time, in order to evaluate the pasted instances and the first in the image The degree of overlap between the original instances is defined by an overlap rate index. :

[0049] Formula 8;

[0050] in Indicates the area of ​​the region. The union of the bounding boxes of all pasted instances; if the occlusion ratio of the original instance is... If the original instance is found to be severely occluded, it will be discarded from the processing set; enhance the image. The corresponding final enhanced pseudo-tag set for:

[0051] Formula 9.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] (1) It specifically addresses the dual challenges of class imbalance and scale imbalance in semi-supervised blood cell detection by adopting a detector-independent design that is compatible with mainstream target detection architectures. It automatically mines high-quality instances with low annotation budgets and performs adaptive data augmentation on underrepresented classes such as minority classes and small scales, overcoming the limitations of existing methods that handle a single imbalance problem alone, and balancing detection performance and annotation efficiency.

[0054] (2) An innovative dual-view high-quality instance selection strategy is proposed. By combining the classification confidence and localization consistency criteria of dual parallel detectors, pseudo-labels are screened, effectively filtering noisy instances and constructing a highly reliable dynamic instance pool. This solves the problem of inconsistent pseudo-label quality in single detectors and provides high-quality data support for model training.

[0055] (3) A background-aware pasting enhancement strategy is proposed, which prioritizes pasting the sampled instances to the background area in the center of the image, quantifies the degree of occlusion and filters severely occluded instances, avoids destroying the structural integrity and semantic context of the original image, overcomes the defect of traditional instance pasting enhancement that easily interferes with the original target detection, and improves the effectiveness of data enhancement. Attached Figure Description

[0056] Figure 1 This is the overall flowchart of the present invention.

[0057] Figure 2 This is a schematic diagram of the before and after images and pseudo-labels after enhancement in an example of the present invention.

[0058] Figure 3 This invention is used to compare the detection results obtained with those obtained by other methods. Detailed Implementation

[0059] In conjunction with the content of this invention, the following embodiments for target detection in blood cell images are provided. These embodiments are implemented on a computer with an Intel(R) Xeon(R) CPU E5-2678v3 @ 2.50GHz CPU, an Nvidia GTX2080Ti GPU, and 32.0GB of memory, and the programming language is Python.

[0060] Step 1: Preparation of training set for blood cell detection and dual-view-based Figure 1 Building a consistent, high-quality instance pool:

[0061] In this embodiment, images containing three types of blood cells—red blood cells, platelets, and white blood cells—were obtained from the BCCD public dataset. This dataset contains 364 images, each 640 pixels in size. 480 blood cell images were divided into training, validation, and test sets, with corresponding image counts of 256, 72, and 36, respectively. To validate the semi-supervised method of this invention, 5% of the samples in the training set were randomly selected to obtain manually labeled blood cell detection data as the gold standard, while the remaining 95% of the samples were treated as unlabeled images.

[0062] The detector structure in this embodiment uses the Faster R-CNN architecture. Two Faster R-CNN detectors with different initialization parameters are set up, denoted as follows: and Unlabeled images Input detectors respectively and Generate the corresponding initial pseudo-tags. and Subsequently, for each cell type ,Will and The subset of the corresponding category in is denoted as and Calculate the intersection-union (IoU) matrix of the bounding boxes of the two subsets. ( , (These represent the number of bounding boxes for the two subsets, respectively). represent The A detection box and The The IoU value of each detection box. Based on Filter out noisy instances by selecting bidirectional matching instance pairs with consistent positioning.

[0063] Formula 1;

[0064] in , , It is the set of instance pairs that match the predictions of the two networks after filtering.

[0065] For each bidirectional matching pair Define a quality score that integrates classification confidence and localization consistency:

[0066] Formula 2;

[0067] in This represents the predicted category probability distribution. Because... and To represent the same instance, assign it the same quality score. .

[0068] Secondly, a quality scoring threshold is set based on this. Only retain those with a quality score higher than [the previous one]. Instances, as a collection of high-quality instances. ::

[0069] Formula 3;

[0070] Next, using Build an instance pool from high-quality instances. The length of this instance pool is fixed. Considering that minority class instances are extremely scarce within each image, to alleviate the class imbalance problem in the instance pool, from each class... One instance is randomly selected from the pool and added to the dynamic instance pool. At the end, A first-in-first-out (FIFO) dynamic update strategy is adopted to maintain only the latest samples to ensure data timeliness.

[0071] Step 2: Class and scale balanced sampling in the instance pool:

[0072] Based on the instance pool composed of the aforementioned high-quality instances, this embodiment randomly samples instance samples from the instance pool, and pastes them onto the original image to generate two different versions of enhanced images, which are then used as... and The input is used to generate the detector. Taking the corresponding enhanced image as an example, firstly, for each batch of images, the category sampling probability is set according to the category frequency. as follows:

[0073] Formula 4;

[0074] Formula 5;

[0075] in Indicates detector Category in the predicted reliable pseudo-labels The reciprocal of the frequency, For set The average area of ​​the middle bounding box. Used to control the degree of contribution of scale information. It is a very small constant to avoid division by zero errors.

[0076] Secondly, based on probability After determining the category to be sampled, randomly select from that category. One instance is used for the paste operation. It is an adaptive number of instances, defined as follows:

[0077] Formula 6;

[0078] in Represents the maximum number of cells contained in any image of the labeled subset. Indicates a single unlabeled image pseudo-tags;

[0079] For each category From the dynamic pool respectively Medium sampling Each image has a sample set. The final sample set for each image is... ,in , These represent the instance image patch and its corresponding local label, respectively.

[0080] Step 3: Enhanced Background-Aware Instance Pasting:

[0081] First, for unlabeled images Gamma correction is performed to enhance the contrast between the foreground and background of the image. Then, an adaptive threshold is calculated using the Otsu method to generate a binary background mask, thereby obtaining the pixel set of the background region. For each in the set Pixels Calculate the minimum Euclidean distance to the foreground-background boundary. A selection probability is defined based on this distance, such that pixels farther from the boundary have a higher selection probability:

[0082] Formula 7;

[0083] Secondly, for instance collections Each instance in Based on probability distribution Randomly sample a pixel Using this instance as the center, paste it to the corresponding position in the image. After the pasting operation is complete, based on pixels... Position recalibration instance The bounding box coordinates, and the calibrated set of instances are denoted as To evaluate the pasted instances and the first in the image The degree of overlap between the original instances is defined by an overlap rate index. :

[0084] Formula 8;

[0085] in Indicates the area of ​​the region. The union of the bounding boxes of all pasted instances; if the occlusion ratio of the original instance is... (in If the original instance is found to be severely occluded, it will be discarded from the processing set. (Image enhancement) The corresponding final enhanced pseudo-tag set for::

[0086] Formula 9;

[0087] Figure 2 This is a comparison of background-aware instance pasting enhancement on an unlabeled image before and after enhancement, and an example image of the enhanced pseudo-label.

[0088] Step 4: Cross-supervised learning based on instance-based augmented data:

[0089] For two detectors and pseudo-tags and Perform steps 2 and 3 independently to obtain augmented data pairs. and . Detector Pseudo-labels are used for monitoring detectors At the same time, the detector Pseudo-labels are used for monitoring detectors Calculate the cross-pseudo-label supervision loss between the two:

[0090] Formula 10;

[0091] Formula 11;

[0092] in, For classifying losses, For bounding box regression loss; detector and Calculate the supervised loss on the labeled data respectively:

[0093] Formula 12;

[0094] The total loss during training is defined as:

[0095] Formula 13;

[0096] in, These are weighting coefficients used to adjust the proportion of the unsupervised loss component in the total loss. After loss calculation, backpropagation is performed to update the network, and training is completed after multiple iterations. In the prediction phase, the trained detector is used. Perform inference on the test images. Figure 3 The detection results obtained after inputting blood cell images into the model trained above are compared with the results of other methods. It can be seen that it is closest to the real label and has good detection capabilities for different types of cells.

Claims

1. An automated blood cell detection method based on semi-supervised learning, the method comprising: Step 1: Building a high-quality instance pool based on dual-view consistency; In a training batch, containing Image of blood cells with annotations and its annotation , Zhang's unlabeled blood cell image For a single labeled image Its labeling information Include One foreground cell instance; One-hot encoding representing category labels, Represents the bounding box coordinates. It is a collection of cell categories; Use two detectors with identical structures but different initialization parameters. and As two different views, they learn through a pseudo-label cross-supervision mechanism on unlabeled images; The quality of pseudo-labels is evaluated using a pseudo-label quality scoring method based on dual-view consistency and a dynamic instance pool construction method, and an instance pool is constructed using high-quality instance-level pseudo-labels. Step 2: Sampling for class and scale balance in the instance pool; Instance samples are randomly sampled from the instance pool and pasted onto the original image to generate two different versions of enhanced images, which are then used as... and Input; For the generation detector Taking the corresponding enhanced image as an example, firstly, for each batch of images, the category sampling probability is set according to the category frequency. as follows: Official 4; Official 5; in This represents a weighted score that combines category frequency with size. Indicates detector Category in the predicted reliable pseudo-labels The reciprocal of the frequency, For set The average area of ​​the middle bounding box; Used to control the degree of contribution of scale information. It is a constant less than a set threshold to avoid division by zero errors; The category sampling probability defined by Formula 4 can increase the sampling probability of rare categories and small-scale categories, thereby alleviating the imbalance problem. Secondly, based on probability After determining the category to be sampled, samples are adaptively and randomly selected from instances of that category for subsequent pasting operations; to ensure that the number of cells in the pasted image is within a reasonable range, a maximum allowed number of instances is introduced. This parameter represents the maximum number of cells contained in any image of the labeled subset; For those with false labels Single unlabeled image The total number of samples that need to be sampled for: Official 6; This represents the predefined maximum number of instances for each category. From the dynamic pool respectively Medium sampling Each image has one sample; the final sampled instance set for each image is... ,in and These represent instance image patches and their corresponding instances, respectively. Step 3: Based on the instance set obtained from the above steps These instances are then pasted into the original image for data augmentation; Step 4: Cross-supervised learning based on instance-based augmented data: For two detectors and pseudo-tags and Perform steps 2 and 3 independently to obtain instance augmentation data pairs. and ; to detector Pseudo-labels are used for monitoring detectors At the same time, the detector Pseudo-labels are used for monitoring detectors Calculate the cross-pseudo-label supervision loss between the two. and : Official 10; Official 11; in, This indicates the number of unlabeled images in the training set. , Indicates detector-based and The predicted pseudo-labels are then processed in steps 2 and 3 to obtain the instance-enhanced image. , This represents the pseudo-tag corresponding to the instance enhancement. For classifying losses, The bounding box regression loss; in addition, the detector and Calculate the supervision loss on the labeled data respectively. : Official 12; The total loss during training is defined as: Official 13; in, The weights are used to adjust the proportion of the unsupervised loss component in the total loss. Based on the above loss function definition, the detector parameters are updated through backpropagation to obtain the trained detection model. and ; Step 5: During the testing phase, a blood cell test image is used. Using a trained detector Prediction is performed to obtain blood cell test results. Furthermore, non-maximum suppression is used to merge detection instances with consistent categories and high overlap (cross-union ratio greater than a threshold) into a single detection instance.

2. The method for automatic blood cell detection based on semi-supervised learning as described in claim 1, characterized in that, Step 1 employs a pseudo-label quality scoring method based on dual-view consistency and a dynamic instance pool construction method to evaluate the quality of pseudo-labels and utilize high-quality instance-level pseudo-labels. The specific method is as follows: First, the unlabeled image Input detectors respectively and Generate the corresponding initial pseudo-tags. and ; Medium category The set of cells is denoted as Its size is recorded as ; Medium category The set of cells is denoted as Its size is recorded as ; for categories The intersection-union ratio (CUI) is calculated for any pair of bounding boxes of two targets in these two subsets, resulting in an CUI matrix. ,in represent The A detection box and The The IoU value of each detection box; based on Filter out noisy instances by selecting bidirectional matching instance pairs with consistent positioning. Official 1; in , ; It is the set of instance pairs that match the predictions of the two networks after filtering; and They represent The and the One detection box, , They represent The The and the first One detection box; For each bidirectional matching pair Define a quality score that integrates classification confidence and localization consistency: Official 2; in Represents the predicted category probability distribution; due to and To represent the same instance, assign it the same quality score. ; Secondly, a quality scoring threshold is set based on this. Only retain those with a quality score higher than [the previous one]. Instances, as a collection of high-quality instances. : Official 3; Next, using Build an instance pool from high-quality instances. The length of this instance pool is fixed. Considering that minority class instances are extremely scarce within each image, to alleviate the class imbalance problem in the instance pool, from each class... One instance is randomly selected from the pool and added to the dynamic instance pool. At the end, A first-in-first-out (FIFO) dynamic update strategy is adopted to maintain only the latest samples to ensure data timeliness.

3. The method for automatic blood cell detection based on semi-supervised learning as described in claim 1, characterized in that, In step 2 .

4. The method for automatic blood cell detection based on semi-supervised learning as described in claim 1, characterized in that, The specific method for step 3 is to use an enhanced background-aware pasting approach, which includes: First, for unlabeled images Gamma correction is performed to enhance the contrast between the foreground and background of the image. An adaptive threshold is calculated using the large-law method to generate a binary background mask, thereby obtaining the pixel set of the background region. For each in the set Pixels Calculate the minimum Euclidean distance to the foreground-background boundary. A selection probability is defined based on this distance, such that pixels farther from the boundary have a higher selection probability: Official 7; Secondly, for instance collections Each instance in Based on probability distribution Randomly sample a pixel Using this instance as the center, paste it to the corresponding position in the image; after the pasting operation is complete, based on the pixel... Position recalibration instance The bounding box coordinates, and the calibrated set of instances are denoted as ; At the same time, in order to evaluate the pasted instances and the first in the image The degree of overlap between the original instances is defined by an overlap rate index. : Official 8; in Indicates the area of ​​the region. The union of the bounding boxes of all pasted instances; if the occlusion ratio of the original instance is... If the original instance is found to be severely occluded, it will be discarded from the processing set; enhance the image. The corresponding final enhanced pseudo-tag set for: Official 9.

Images