An instance segmentation network optimization method, device and medium

Abstract

The application relates to an instance segmentation network optimization method, device and medium, the method comprising: obtaining a batch of tail data samples, performing data augmentation on the tail data samples to obtain tail extension samples; performing contrast learning on the tail data samples and the tail extension samples to obtain a tail contrast loss; constructing pseudo semantic labels of the tail extension samples; obtaining a batch of head data samples, the head data samples and the tail data samples both comprising true value labels; inputting the head data samples, the tail data samples and the tail extension samples into a preset instance segmentation network, calculating a segmentation mask loss and a category loss based on the true value labels and the pseudo semantic labels; adjusting the preset instance segmentation network based on the tail contrast loss, the segmentation mask loss and the category loss to obtain an optimized instance segmentation network. The application solves the problem that the recognition accuracy of existing models for tail rare categories is low, resulting in poor long-tail instance segmentation effect.

Description

Technical Field

[0001] The present application relates to the technical field of image processing model optimization, and in particular to an instance segmentation network optimization method, device, and medium. Background Art

[0002] Instance segmentation aims to associate each pixel in an image with its corresponding semantic category and instance. It combines the characteristics of semantic segmentation, which requires pixel-level classification, with the characteristics of target detection, which requires locating different instances. Instance segmentation can provide computers with more fine-grained scene information for image understanding and achieve more accurate object detection, and has always been an important research direction in image processing. In practical applications, most data is not evenly distributed, but rather skewed, with a large number of instances in most common categories and a very small proportion of instances in rare categories, resulting in a long tail overall. This long-tail distribution causes the model to be more inclined to learn and optimize the features of the head categories during training, while its ability to recognize the tail categories is insufficient, thus affecting the accuracy of instance segmentation. Therefore, instance segmentation methods targeting the long-tail distribution of data are becoming an increasingly hot research topic.

[0003] The core of instance segmentation tasks targeting long-tail data distribution lies in addressing the imbalanced data distribution. Existing research methods can be broadly categorized into two main categories: class-balanced sampling and cost-sensitive learning. Class-balanced sampling methods mitigate the impact of long-tail distributions by adjusting the selection strategy for training samples. Common approaches include oversampling and undersampling. Oversampling increases the number of tail class samples by duplicating existing samples or using simple data augmentation techniques such as cropping and flipping to balance the data distribution. Undersampling reduces the number of head class samples to achieve a relatively balanced representation across classes during training. While these methods can achieve a superficial balance in terms of sample class, they essentially require the model to repeatedly learn from a small amount of data from tail class samples. This lacks sufficient sample variance and can easily lead to overfitting and poor generalization. Cost-sensitive learning mitigates the long-tail problem by assigning different training weights to samples from different classes. However, determining the weights for each class is challenging, especially in real-world scenarios where the correlations between classes and the weight distribution may not be obvious. Improper weighting can lead to the model learning erroneous information during training, impacting performance. During training, the model assigns higher loss weights to rare categories, causing errors in tail categories to be penalized more severely. This approach does not alter the data distribution, but rather encourages the model to focus more on rare categories by modifying the loss function. However, the weight distribution for each category in the loss function and the associations between categories may not be obvious. Modifying the loss function alone may cause the model to learn incorrect information during training, impacting performance.

[0004] Therefore, in existing technologies, class-balanced sampling methods lack sufficient sample variance, which can easily lead to model overfitting and poor generalization. Cost-sensitive learning methods may cause the model to learn incorrect information. Models trained with both methods have low recognition accuracy for rare tail classes, resulting in poor long-tail instance segmentation performance. Therefore, an instance segmentation network that can solve the long-tail instance segmentation problem is urgently needed. Summary of the Invention

[0005] In order to overcome the problem that the existing models have low recognition accuracy for rare categories at the tail, resulting in poor long-tail instance segmentation, the present application provides an instance segmentation network optimization method, device and medium.

[0006] In order to solve the above technical problems, the present application provides an instance segmentation network optimization method, including:

[0007] Obtain batches of tail data samples, and perform data augmentation on the tail data samples to obtain tail extended samples;

[0008] Perform comparative learning on the tail data samples and the tail extension samples to obtain the tail contrast loss;

[0009] Construct pseudo semantic labels for tail extension samples;

[0010] Get batches of header data samples, both header and tail data samples contain true value labels;

[0011] Input the head data sample, tail data sample and tail extension sample into the preset instance segmentation network, and calculate the segmentation mask loss and category loss based on the true value label and pseudo semantic label;

[0012] Based on tail contrast loss, segmentation mask loss and category loss, the preset instance segmentation network is adjusted to obtain the optimized instance segmentation network.

[0013] In a second aspect, the present application also provides an instance segmentation network optimization system, comprising:

[0014] The data augmentation module is used to obtain batches of tail data samples and perform data augmentation on the tail data samples to obtain tail extended samples;

[0015] The comparison module is used to perform comparative learning on the tail data samples and the tail extension samples to obtain the tail comparison loss;

[0016] Construction module, used to construct pseudo semantic labels for tail extension samples;

[0017] The acquisition module is used to obtain batches of head data samples, where both the head data samples and the tail data samples contain true value labels;

[0018] A calculation module is used to input the head data sample, the tail data sample and the tail extension sample into the preset instance segmentation network, and calculate the segmentation mask loss and the category loss based on the true value label and the pseudo semantic label;

[0019] The optimization module is used to adjust the preset instance segmentation network based on tail contrast loss, segmentation mask loss and category loss to obtain an optimized instance segmentation network.

[0020] In a third aspect, the present application also provides a computing device, including a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, the steps of the instance segmentation network optimization method as described above are implemented.

[0021] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores instructions. When the instructions are executed on a terminal device, the terminal device executes the steps of an instance segmentation network optimization method.

[0022] The beneficial effects of the present application are: by performing data augmentation on a batch of tail data samples, tail extension samples are obtained, which can increase the sample diversity of the tail category and provide more tail category feature variants for subsequent model training to preliminarily balance the data distribution for training. The tail extension samples are compared with the tail data samples to obtain the tail contrast loss. The tail contrast loss can be used to understand the network's ability to recognize inter-class discriminative features and the network's ability to focus on intra-class invariant features. Then, pseudo-semantic labels for the tail extension samples are constructed, and both the head data samples and the tail data samples contain true value labels. The batch of head data samples, tail data samples and tail extension samples are input into the preset instance segmentation network. The segmentation mask loss and category loss are calculated based on the true value labels and pseudo-semantic labels. The difference between the label and the predicted value of the tail input value can be understood through the segmentation mask loss and category loss. Based on the tail contrast loss, segmentation mask loss and category loss, the preset instance segmentation network is adjusted to achieve network optimization. This can not only improve the network's ability to recognize the inter-class discriminative features of long-tail data and increase the network's attention to the intra-class invariant features of long-tail data, but also reduce the difference between the label of the input sample and the corresponding predicted value output by the network, thereby improving the recognition accuracy of the optimized instance segmentation network for long-tail data, and thus improving the segmentation effect of the network model on long-tail instances. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of an example segmentation network optimization method according to an exemplary embodiment of the present application is shown;

[0024] Figure 2A distribution histogram of the number of instances of head data and long-tail data in the prior art;

[0025] Figure 3 A network structure diagram for implementing an instance segmentation network optimization method in an exemplary embodiment of the present application;

[0026] Figure 4 This is a schematic diagram of a scene image of a tail data sample in an exemplary embodiment of the present application;

[0027] Figure 5 This is a schematic diagram of a predicted image obtained by processing a tail data sample through an optimized instance segmentation network in an exemplary embodiment of the present application;

[0028] Figure 6 The diagram is a structural diagram of an instance segmentation network optimization system according to an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0029] The following examples are provided to further explain and supplement the present application and do not constitute any limitation to the present application.

[0030] The following describes an example segmentation network optimization method, device, and medium according to an embodiment of the present application in conjunction with the accompanying drawings.

[0031] An embodiment of the present application provides an instance segmentation network optimization method, which is applied to a terminal device. In the present application scheme, the terminal device is used as the execution subject to illustrate the present application scheme, and the terminal device is used to execute the steps of an instance segmentation network optimization method.

[0032] See also Figure 1 , Figure 1 An example segmentation network optimization method is shown as an exemplary embodiment of the present application. Figure 1 As shown, the present application provides an instance segmentation network optimization method, including:

[0033] Step S11, obtaining a batch of tail data samples, and performing data augmentation on the tail data samples to obtain tail extended samples;

[0034] Step S12, performing comparative learning on the tail data sample and the tail extension sample to obtain the tail comparative loss;

[0035] Step S13, constructing a pseudo semantic label of the tail extension sample;

[0036] Step S14, obtaining a batch of header data samples, where both the header data samples and the tail data samples contain true value labels;

[0037] Step S15, input the head data sample, tail data sample and tail extension sample into the preset instance segmentation network, calculate the segmentation mask loss and the class loss based on the true value label and the pseudo semantic label;

[0038] Step S16, adjust the preset instance segmentation network based on the tail contrast loss, the segmentation mask loss and the class loss, to obtain an optimized instance segmentation network.

[0039] The instance segmentation network optimization method of the embodiment can increase the sample diversity of the tail classes by data augmentation on the batch of tail data samples to obtain tail extension samples, provide more tail class feature variants for subsequent model training, and preliminarily balance the data distribution used for training. The tail extension samples and the tail data samples are compared and learned to obtain the tail contrast loss, which can understand the recognition ability of the network to inter-class discriminative features and the attention ability of the network to intra-class invariance features. Then, the pseudo semantic label of the tail extension sample is constructed, the head data sample and the tail data sample both contain the true value label, and the batch of head data samples, tail data samples and tail extension samples are input into the preset instance segmentation network. The segmentation mask loss and the class loss are calculated based on the true value label and the pseudo semantic label. The difference between the label and the predicted value of the tail input value can be understood through the segmentation mask loss and the class loss. The preset instance segmentation network is adjusted based on the tail contrast loss, the segmentation mask loss and the class loss, to realize network optimization. Not only can the recognition ability of the network to inter-class discriminative features of long-tail data and the attention of the network to intra-class invariance features of long-tail data be improved, but also the difference between the label of the input sample and the corresponding predicted value output by the network can be reduced, so that the recognition accuracy of the optimized instance segmentation network obtained by optimization on long-tail data can be improved, and the segmentation effect of the network model on long-tail instances can be improved. The number of tail data samples is small, the number of tail extension samples is large, and the number of tail data samples and the number of tail extension samples are balanced with the number of head data samples.

[0040] The instance quantity of the head data and the long-tail data in the prior art is as shown in Figure 2 Figure 2 ​In the example, frequent represents a frequent category, common represents a common category, and rare represents a rare category. The sum of the number of categories of frequent and common is used as the number of instances of the head data sample, and rare is used as the number of instances of the tail data sample. In this embodiment, the tail data sample is expanded by generative AI technology to achieve data augmentation of the tail data sample. Although traditional category-balanced sampling methods (such as oversampling and undersampling) can solve the problem of unbalanced sample quantity, they cannot fundamentally solve the problem of the diversity of learnable features of the network model due to the lack of sufficient sample variance. Therefore, this embodiment combines generative AI technology to generate tail extension samples with diversity, aiming to balance the data distribution while allowing the model to be exposed to more tail category variants, so as to improve the network model's feature recognition ability for the tail category. At the same time, by performing comparative learning on the tail data sample and the tail extension sample, comparing the similarities and differences between the two samples, and realizing comparative learning, the preset instance segmentation network can learn inter-class difference features and intra-class invariant features, strengthen the feature association between the two samples, and thus improve the network model's recognition ability for the tail category. At the same time, it can also improve the generalization ability of the network model and its robustness to noise and interference, making the network model more robust.

[0041] Therefore, this embodiment performs data augmentation on the tail data samples and uses the contrastive learning method to strengthen the feature association between a small number of tail data samples and a large number of augmented tail extension samples. This can balance the distribution ratio of head category samples and tail category samples input into the network model during training, and guide the network model to learn the consistent features of a large number of tail extension samples, thereby improving the segmentation performance of the model on the tail category and improving the segmentation effect of long-tail instances.

[0042] Optionally, a contrastive learning is performed on the tail data samples and the tail extension samples to obtain the tail contrast loss, including:

[0043] Perform feature extraction on the tail data samples to obtain tail data features;

[0044] Extract features of the tail extension samples to obtain tail extension data features;

[0045] Calculating the first cosine similarity between each tail extension data feature and each tail data feature to obtain a similarity set corresponding to each tail extension data feature;

[0046] The calculation formula for the first cosine similarity is as follows:

[0047]

[0048] Where sim(u, v) represents the first cosine similarity, u represents the tail extension data feature, v represents the tail data feature, and T represents transposition;

[0049] Based on the tail extended data features and the similarity set, the feature contrast loss of the tail extended data features is obtained;

[0050] Based on the tail extension data features and the similarity set, the tail extension data features are feature enhanced, and the enhancement contrast loss of the tail extension data features is determined;

[0051] The tail contrast loss is formed based on the feature contrast loss and the enhancement contrast loss.

[0052] In this embodiment, feature extraction is performed on the tail data samples and the tail extension samples respectively to obtain tail data features and tail extension data features, and the first cosine similarity between each tail extension data feature and each tail data feature is calculated to obtain a similarity set corresponding to each tail extension data feature. Then, based on the tail extension data features and the similarity set, the feature contrast loss and the enhancement contrast loss of the tail extension data features are obtained to form the tail contrast loss. Since the feature contrast loss can reflect the network model's ability to recognize inter-class discriminative features, and the enhancement contrast loss can reflect the network model's ability to pay attention to intra-class invariant features, this embodiment can improve the network model's ability to recognize inter-class discriminative features and enhance the network model's attention to intra-class invariant features during subsequent training of the preset instance segmentation network through the tail contrast loss, thereby improving the recognition accuracy of the optimized instance segmentation network obtained by subsequent optimization training for long-tail data, and further improving the segmentation effect of the network model on long-tail instances.

[0053] In an exemplary embodiment, when feature extraction is performed on the tail data samples and the tail extension samples, feature extraction is also performed on the head data samples, and the three types of samples are input into a three-branch parallel weight-sharing feature extraction network to extract the corresponding feature representations. The weight-sharing feature extraction network used in this embodiment simplifies the structure of the network model by sharing weights, reduces the total number of parameters of the model, and reduces storage requirements and computational complexity. At the same time, weight sharing can ensure the consistency of the extraction conditions of feature extraction, thereby helping the model establish better feature connections between samples of different categories, thereby improving the ability to distinguish rare categories.

[0054] At the same time, batches of tail extension data features can be embedded into the constructed query Q, with each tail extension data feature serving as a query feature q in the query Q. The tail data features are then used as keys K, with each tail data feature serving as a key feature k in the key K. For each query feature q, the cosine similarity between it and each key feature k is calculated, and this cosine similarity is used as the first cosine similarity between the corresponding tail extension data feature and a tail data feature.

[0055] Optionally, based on the tail extension data feature and the similarity set, a feature contrast loss of the tail extension data feature is obtained, including:

[0056] Taking the multiple similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the multiple similarity sets except the to-be-compared sets;

[0057] The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples;

[0058] The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples;

[0059] The feature contrast loss is obtained based on the tail extended data features corresponding to the positive sample, negative sample and the set to be compared;

[0060] The formula for calculating feature contrast loss is as follows:

[0061]

[0062]

[0063] in, represents the feature contrast loss of the tail extension data feature corresponding to the current set to be compared, q represents the tail extension data feature corresponding to the current set to be compared, q + Represents the positive sample of the tail extension data feature corresponding to the current set to be compared, q - Represents the negative sample of the tail extended data feature corresponding to the current set to be compared, τ represents the temperature coefficient, which is an adjustable parameter used to affect the model's discrimination of samples, λ represents the weight of measuring the similarity between positive samples and negative samples, and the weight is used to reduce the influence of negative samples similar to positive samples. neg Indicates the number of negative samples.

[0064] In this embodiment, by using a similarity set consisting of the first cosine similarity between the tail extension data feature and the tail data feature, positive samples and negative samples corresponding to each tail extension data feature are constructed, a comparison relationship between the two samples is established, and corresponding feature comparison losses are obtained based on the positive samples and negative samples of each tail extension data feature. This can effectively guide the model to capture the inter-class discriminative features between the tail extension data feature and the tail data feature, so as to improve the model's recognition ability of the inter-class discriminative features, thereby strengthening the feature association between the tail extension sample and the tail data sample, and enhancing the model's recognition ability for different categories without relying on complex weight settings.

[0065] In this embodiment, for each query feature q of the query Q constructed by the tail extended data feature, its similar features (i.e., the tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity) and itself are designated as positive samples according to the corresponding cosine similarity calculation result, and other query features are designated as negative samples. The contrast loss is calculated between the positive and negative samples, so that the model shortens the distance between similar samples and expands the distance between different samples in the feature space, so as to learn more accurate category boundaries.

[0066] Optionally, based on the tail extension data feature and the similarity set, the tail extension data feature is enhanced, and the enhanced contrast loss of the tail extension data feature is determined, including:

[0067] Taking the multiple similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the multiple similarity sets except the to-be-compared sets;

[0068] The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples;

[0069] The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples;

[0070] The first cosine similarity greater than the threshold in the set to be compared is used as the second target similarity;

[0071] Performing a feature aggregation operation on the tail data features corresponding to the second target similarity and the tail extended data features corresponding to the set to be compared to obtain enhanced features;

[0072] Based on the positive samples, negative samples and enhanced features, the enhanced contrast loss of the tail extended data features corresponding to the set to be compared is obtained;

[0073] The calculation formula of enhanced contrast loss is as follows:

[0074]

[0075] Among them, q represents the enhanced feature corresponding to the current set to be compared, q + Represents the positive sample of the tail extension data feature corresponding to the current set to be compared, q - It represents the negative sample of the tail extended data feature corresponding to the current set to be compared. τ represents the temperature coefficient, which is an adjustable parameter used to affect the model's discrimination of samples.

[0076] In this embodiment, a similarity set consisting of the first cosine similarity between the tail extension data features and the tail data features is used to construct positive and negative samples corresponding to each tail extension data feature, establish a comparative relationship between the two samples, and determine a second target similarity from the similarity set. Based on the second target similarity, an enhanced feature corresponding to each tail extension data feature is obtained. An enhanced contrast loss is then derived for each tail extension data feature based on the positive samples, negative samples, and enhanced features. This effectively guides the model to capture the intra-class invariant features between the tail extension data features and the tail data features, thereby improving the model's ability to recognize intra-class invariant features. This further strengthens the feature association between the tail extension samples and the tail data samples, and further enhances the model's ability to recognize different classes without relying on complex weight settings.

[0077] Optionally, construct pseudo semantic labels for the tail extension samples, including:

[0078] Perform feature extraction on the tail extension samples to obtain dense features;

[0079] Downsample the dense features to obtain reduced-dimensional features, and use the reduced-dimensional features as pseudo-category labels for the tail-extended samples;

[0080] Calculate the second cosine similarity between each reduced feature and each dense feature;

[0081] Based on the second cosine similarity and dense features, the score mapping of the tail extension samples is obtained;

[0082] Calculate the pseudo mask of the corresponding tail extension sample based on the score mapping;

[0083] Pseudo semantic labels are formed based on pseudo masks and pseudo category labels.

[0084] In this embodiment, dense features of the tail-extended samples are extracted and downsampled, and the resulting reduced-dimensionality features are used as pseudo-category labels for the tail-extended samples. Based on the second cosine similarity between each reduced-dimensionality feature and each dense feature, a score mapping of the tail-extended samples is obtained. A pseudo-mask of the corresponding tail-extended sample is calculated based on the score mapping, and a pseudo-semantic label is formed based on the pseudo-mask and pseudo-category label. This pseudo-semantic label is then used to optimize the pre-set instance segmentation network, thereby improving the recognition accuracy of the optimized instance segmentation network for long-tail data and enhancing the segmentation performance of the network model for long-tail instances.

[0085] In an exemplary embodiment, the pre-trained backbone network is used to extract features from the tail extension samples to obtain dense features I∈R H×W×E , H×W×E represents the size of the feature space. Then construct the query Q and key K: take the dense feature I as a key K in the set of keys K. i,j , perform bilinear downsampling on I to form a reduced dimension feature, and use the reduced dimension feature as a query vector q∈R in the set of query Q E , where the query Q∈R H×W×E , H′×W×E′ represents the size of the down-sampling space. For each query vector q in the query Q, calculate its correlation with each key K in the key K i，j The cosine similarity between the two is used as the second cosine similarity between the corresponding dimensionality reduction feature and a dense feature. Based on the second cosine similarity and the dense feature, the score mapping S∈R of the tail extension sample is obtained H×W×N .

[0086] The calculation formula of the score mapping is as follows:

[0087] S i，j，q =sim(Q q , K i，j );

[0088] Among them, S i，j，q Represents the score mapping attention map composed of the second cosine similarity corresponding to the qth query vector, which reflects the relationship between the qth query vector and the key K i，j The second cosine similarity between each position in the corresponding H×W×E feature map, Q q represents the qth query vector, K i，j represents the bond at position (i, j) in the H′×W′×E′ space.

[0089] Optionally, a pseudo mask of the corresponding tail extension sample is calculated based on the score mapping, including:

[0090] Normalize the score map to obtain the initial pseudo mask;

[0091] Obtain the mask score of the initial pseudo mask, and use the mask score as the confidence score of the initial pseudo mask;

[0092] A non-maximum suppression operation is performed on the initial pseudo mask based on the confidence score to obtain a pseudo mask of the tail extension sample corresponding to the score mapping.

[0093] In this embodiment, the score mapping is normalized, and different score mappings can be standardized to obtain initial pseudo-masks of the same format. The mask score of the initial pseudo-mask is obtained, and the mask score is used as the confidence score of the initial pseudo-mask. Based on the confidence score, a non-maximum suppression operation is performed on the initial pseudo-mask to remove redundant masks in the initial pseudo-mask. A pseudo-mask that meets the accuracy requirements for the tail extension samples corresponding to the score mapping is obtained. This facilitates reducing the interference of redundant masks during subsequent model training, thereby improving the accuracy of model training and further improving the segmentation effect of the trained network model on long-tail instances.

[0094] In an exemplary embodiment, the score map is normalized to the range [0, 1] to generate an initial pseudo-mask. The initial pseudo-mask is further used to calculate its mask score as a confidence score. A threshold τ is set to convert the initial pseudo-mask into a binary mask, and the binary masks are sorted according to the confidence score. Redundant masks are removed through non-maximum suppression to obtain the pseudo-mask of the tail extension sample corresponding to the score map. The calculation formula is as follows:

[0095] M=NMS(Maskness(Norm(S i，j，q )));

[0096] Among them, M represents a pseudo mask, NMS represents a non-maximum suppression operation, Maskness represents an operation for calculating the confidence score of the initial pseudo mask, and Norm represents a normalization operation on the score mapping.

[0097] The confidence score of the initial pseudo mask is calculated as follows:

[0098]

[0099] Among them, maskness represents the confidence score of the initial pseudo mask, N f represents the initial pseudo mask P i The number of foreground pixels represents the initial pseudo mask P i The number of pixels with eigenvalues ​​greater than τ, P i represents the i-th initial pseudo mask.

[0100] Optionally, the true value label includes a true mask and a true category label; the head data sample, the tail data sample, and the tail extension sample are input into a preset instance segmentation network, and the segmentation mask loss and the category loss are calculated based on the true value label and the pseudo semantic label, including:

[0101] Inputting the head data sample, the tail data sample, and the tail extension sample into a preset instance segmentation network to obtain the predicted mask and predicted semantic category corresponding to the head data sample, the tail data sample, and the tail extension sample;

[0102] Based on the real mask, pseudo mask and predicted mask, the segmentation mask loss is obtained;

[0103] Based on the true category label, pseudo category label and predicted semantic category, the category loss is obtained.

[0104] In this embodiment, head data samples, tail data samples, and tail extension samples are input into a preset instance segmentation network for training. The resulting model output includes predicted masks and predicted semantic categories corresponding to the head data samples, tail data samples, and tail extension samples. Based on the true masks of the head data samples, tail data samples, pseudo masks of the tail extension samples, and the predicted masks of the three, a segmentation mask loss is obtained to guide the model in segmenting the contour boundaries of the target object. Furthermore, based on the true category labels of the head data samples, tail data samples, pseudo category labels of the tail extension samples, and the predicted semantic categories of the three, a category loss is obtained to enhance the model's ability to understand object categories. This facilitates subsequent adjustment and optimization of the preset instance segmentation network based on the segmentation mask loss and category loss, thereby improving the network model's accuracy in segmenting the contour boundaries of objects in the image and improving the network model's ability to understand the categories of long-tail objects. This improves the recognition accuracy of the optimized instance segmentation network for long-tail data, and further improves the network model's segmentation effect on long-tail instances.

[0105] In an exemplary embodiment, the segmentation mask loss is calculated as follows:

[0106] L mask =γL avg-proj +L max-proj +L pairwise ;

[0107] L avg-proj =L(avg x (m),avg x (m * ))+L(avg y (m),avg y (m * ));

[0108] Lmax-proj = L(max x (m), max x (m * )) + L(max y (m), max y (m * )) ;

[0109]

[0110] where L mask denotes the segmentation mask loss, γ denotes a trade-off hyperparameter, L avg-proj denotes the average projection loss, L denotes the Dice (image segmentation task) loss, m * denotes the predicted mask, m denotes the true mask and the pseudo mask, L avg-proj projects the predicted mask onto the x-axis by performing an avg operation on each axis, projects the true mask and the pseudo mask onto the y-axis; L max-proj denotes the max projection loss, projects the predicted mask onto the x-axis by performing a max operation on each axis, projects the true mask and the pseudo mask onto the y-axis, L avg-proj acts jointly with L max-proj to minimize the difference between the projection of m * and the projection of m; L pairwise denotes the pairwise affinity loss, which supervises the predicted mask in a pixel-wise manner, ensuring label consistency between neighboring pixels; E in denotes the set of pixel edges; y e ∈ {0, 1} is the label of edge e, y e = 1 indicates that the two pixels connected by the edge have the same label, y e = 0 indicates that the two pixels connected by the edge have different labels.

[0111] The formula for calculating the class loss is as follows:

[0112] L cate = L focal + μL sem ;

[0113] L focal = -α t (1 - p t ) λ log(p t ) ;

[0114]

[0115] where L cate denotes the class loss, L focalrepresents the focal loss, used for foreground and background binary classification of images, μ represents a trade-off hyperparameter, L sem represents the negative cosine similarity function, which measures the cosine value of the angle between two vectors and takes its opposite number, L sem Optimize the recognition accuracy of the model for object semantic embedding by minimizing the negative cosine similarity; α t represents a balance factor, used to adjust the loss weight between positive class (foreground) and negative class (background), helping the model better handle the problem of class imbalance, p t represents the predicted probability of the model that the sample belongs to a certain class (foreground or background), λ represents a regulation factor, used to reduce the loss of easy-to-classify samples, so that the model pays more attention to difficult-to-classify samples during training, q * represents the predicted semantic class, q represents the true class label and the pseudo class label.

[0116] Optionally, based on the tail contrast loss, the segmentation mask loss and the class loss, the preset instance segmentation network is adjusted to obtain an optimized instance segmentation network, comprising:

[0117] When the tail contrast loss, the segmentation mask loss and the class loss meet the preset loss requirement, the preset instance segmentation network is used as the optimized instance segmentation network;

[0118] Otherwise, the network parameters in the preset instance segmentation network are adjusted to obtain an adjusted instance segmentation network;

[0119] The head data sample, the tail data sample and the tail extension sample are input into the adjusted instance segmentation network until the corresponding tail contrast loss, segmentation mask loss and class loss meet the preset loss requirement, and the optimized instance segmentation network is obtained.

[0120] In this embodiment, the preset instance segmentation network is adjusted by using the tail contrast loss, the segmentation mask loss and the class loss to measure the gap between the predicted value and the true label until the gap between the predicted value and the true label is reduced to tend to be flat and cannot be significantly reduced, so that the tail contrast loss, the segmentation mask loss and the class loss of the network meet the preset loss requirement, the network is optimized, and the recognition accuracy of the optimized instance segmentation network obtained by optimization can meet the accuracy requirement, thereby significantly improving the segmentation effect of the network model on long-tail instances.

[0121] Please refer to Figure 3 , Figure 3 In an exemplary embodiment of the present application, a network structure diagram for implementing an instance segmentation network optimization method is implemented; as Figure 3As shown, the network structure diagram includes a sample module, a feature extraction module, a contrastive learning module, a pseudo-label pseudo-mask generation module, a segmentation head, and an optimization module. Among them, the sample module includes tail data samples and head data samples that are samples labeled with true value labels (true masks and true category labels), as well as unlabeled tail extension samples obtained by data augmentation of the labeled tail data samples. When implementing an instance segmentation network optimization method, first, the labeled tail data samples, unlabeled tail extension samples, and labeled head data samples are input into the pre-trained weight-sharing backbone network in the feature extraction module for feature extraction to obtain the corresponding tail data features, tail extension data features, and head data features. Secondly, the tail data features and tail extension data input features are embedded in the contrastive learning module, and contrastive learning is performed on the tail data features and tail extension data features to obtain the tail contrast loss. Thirdly, the unlabeled tail extension samples are input into the pseudo-label generation module to construct pseudo semantic labels (pseudo masks and pseudo category labels) for the tail extension samples. The segmentation head then adopts a dual-branch structure consisting of mask prediction and semantic category prediction, with the semantic category prediction branch indicating semantic category probabilities. The segmentation head predicts a mask for each sample, and a parallel branch consisting of a single convolutional layer is added to the final layer of the segmentation head to predict the semantic category of each sample. The head, tail, and extended tail samples are then fed into the segmentation head's pre-set instance segmentation network, which predicts the predicted masks and semantic categories for the head and tail data samples, as well as the predicted masks and semantic categories for the extended tail samples. Finally, the optimization module obtains the ground-truth masks and true category labels from the annotations of the head and tail data samples. The segmentation mask loss and category loss are calculated using the ground-truth masks, pseudo-masks, predicted masks, true category labels, pseudo-category labels, and predicted semantic categories. This effectively utilizes limited annotation resources and significantly reduces data annotation costs. The pre-set instance segmentation network is then adjusted based on the tail contrast loss, segmentation mask loss, and category loss to obtain an optimized instance segmentation network.

[0122] In an exemplary embodiment, batches of head data samples and batches of tail data samples are obtained from an open source long-tail data sample set. Figure 4 is a tail data sample in the open source long-tail data sample set, and the optimized instance segmentation network is used to segment the Figure 4 The tail data sample shown in the figure is predicted, and the predicted image is as follows Figure 5 shown. Figure 5 The predicted category and the confidence level of the category are shown in . The higher the confidence level, the more confident the optimized instance segmentation network is about the result, and the higher the accuracy.

[0123] The present application discloses an instance segmentation network optimization method, which first augments the tail data samples into diverse tail extension samples by combining generative AI technology to preliminarily balance the sample distribution, and then sends the samples to a weight-shared feature extraction network to extract deep features; then, calculates the cosine similarity of the tail data features and the tail extension data features, aggregates the similarity features between the tail data features and the tail extension data features to form a tail enhancement feature, and determines the positive and negative samples of the tail extension data features based on the cosine similarity calculation results; calculates the feature contrast loss between the tail extension data features and the positive and negative samples to achieve feature association between the tail extension samples and the tail data samples; calculates the enhanced contrast loss between the tail enhancement features and the positive and negative samples to achieve feature clustering of the tail data samples and the tail extension samples, thereby improving the network's ability to recognize the inter-class discriminative features of the long-tail data and increasing the network's attention to the intra-class invariant features of the long-tail data. Finally, the tail extension samples are sent to the zero-supervised pseudo-label generation network to generate corresponding pseudo masks and pseudo category labels, and are input into the preset instance segmentation network together with the head data samples and tail data samples for joint training to obtain the segmentation mask loss and category loss. The preset instance segmentation network is optimized and adjusted using the tail contrast loss, segmentation mask loss and category loss to obtain the optimized instance segmentation network.

[0124] This application combines the advantages of contrastive learning to design a feature embedding contrastive learning module. By establishing contrast relationships between different samples, the model is able to effectively capture the similarity features within the tail category and the difference features between different categories, improve the recognition ability of the tail category, enhance the adaptability to long-tail distribution data, and thus improve the overall segmentation performance of the model when the data is long-tail distributed. At the same time, this application introduces a zero-supervised pseudo-label generation network to design a semi-supervised mechanism that combines full supervision, zero supervision and weak supervision. This not only improves the utilization efficiency of samples and reduces the dependence on a large amount of high-quality labeled data, thereby significantly reducing the cost of data labeling, but also can more effectively improve the learning effect and generalization ability of the model.

[0125] See also Figure 6 , Figure 6 An example segmentation network optimization system is shown as an exemplary embodiment of the present application. Figure 6 As shown, the present application provides an instance segmentation network optimization system, including:

[0126] The data augmentation module is used to obtain batches of tail data samples and perform data augmentation on the tail data samples to obtain tail extended samples;

[0127] The comparison module is used to perform comparative learning on the tail data samples and the tail extension samples to obtain the tail comparison loss;

[0128] The constructing module is configured to construct a pseudo mask and a pseudo class label of the tail extension sample, and the head data sample and the tail data sample both contain true value labels;

[0129] The obtaining module is configured to obtain a batch of head data samples, and the head data sample and the tail data sample both contain true value labels;

[0130] The computing module is configured to input the head data sample, the tail data sample and the tail extension sample into a preset instance segmentation network, and calculate a segmentation mask loss and a class loss based on the true value labels and the pseudo semantic labels;

[0131] The optimization module is configured to adjust the preset instance segmentation network based on the tail contrast loss, the segmentation mask loss and the class loss, to obtain an optimized instance segmentation network.

[0132] An instance segmentation network optimization system according to an embodiment of the present application can obtain tail extension samples by data augmentation of a batch of tail data samples through a data augmentation module, can increase sample diversity of tail classes, can provide more tail class feature variants for subsequent model training, and can preliminarily balance data distribution for training. The tail contrast loss can be obtained by contrast learning of the tail extension sample and the tail data sample through a contrast module, the recognition ability of the network to inter-class discriminative features and the attention ability of the network to intra-class invariance features can be understood through the tail contrast loss. The pseudo semantic labels of the tail extension sample can be constructed through a constructing module, the head data sample and the tail data sample both contain true value labels, and the batch of head data samples, tail data samples and tail extension samples can be input into a preset instance segmentation network through an obtaining module, the segmentation mask loss and the class loss can be calculated based on the true value labels and the pseudo semantic labels, and the difference between the labels of the tail input values and the predicted values can be understood through the segmentation mask loss and the class loss. The preset instance segmentation network can be adjusted based on the tail contrast loss, the segmentation mask loss and the class loss through an optimization module, the network optimization can be realized, the recognition ability of the network to inter-class discriminative features of long-tail data and the attention of the network to intra-class invariance features of long-tail data can be improved, the difference between the labels of the input samples and the corresponding predicted values output by the network can be reduced, the recognition accuracy of the optimized instance segmentation network obtained through optimization to long-tail data can be improved, and the segmentation effect of the network model to long-tail instances can be improved.

[0133] Optionally, the contrast module is specifically configured to:

[0134] extract features of the tail data sample to obtain tail data features;

[0135] extract features of the tail extension sample to obtain tail extension data features;

[0136] Calculating the first cosine similarity between each tail extension data feature and each tail data feature to obtain a similarity set corresponding to each tail extension data feature;

[0137] Based on the tail extended data features and the similarity set, the feature contrast loss of the tail extended data features is obtained;

[0138] Based on the tail extension data features and the similarity set, the tail extension data features are feature enhanced, and the enhancement contrast loss of the tail extension data features is determined;

[0139] The tail contrast loss is formed based on the feature contrast loss and the enhancement contrast loss.

[0140] Optionally, the comparison module is specifically configured to:

[0141] Taking the multiple similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the multiple similarity sets except the to-be-compared sets;

[0142] The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples;

[0143] The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples;

[0144] The feature contrast loss is obtained based on the tail extended data features corresponding to the positive samples, negative samples and the set to be compared.

[0145] Optionally, the comparison module is specifically configured to:

[0146] Taking the multiple similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the multiple similarity sets except the to-be-compared sets;

[0147] The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples;

[0148] The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples;

[0149] The first cosine similarity greater than the threshold in the set to be compared is used as the second target similarity;

[0150] Performing a feature aggregation operation on the tail data features corresponding to the second target similarity and the tail extended data features corresponding to the set to be compared to obtain enhanced features;

[0151] Based on the positive samples, negative samples and enhanced features, the enhanced contrast loss of the tail extended data features corresponding to the set to be compared is obtained.

[0152] Optionally, a building block is provided, specifically for:

[0153] Perform feature extraction on the tail extension samples to obtain dense features;

[0154] Downsample the dense features to obtain reduced-dimensional features, and use the reduced-dimensional features as pseudo-category labels for the tail-extended samples;

[0155] Calculate the second cosine similarity between each reduced feature and each dense feature;

[0156] Based on the second cosine similarity and dense features, the score mapping of the tail extension samples is obtained;

[0157] Calculate the pseudo mask of the corresponding tail extension sample based on the score mapping;

[0158] Pseudo semantic labels are formed based on pseudo masks and pseudo category labels.

[0159] Optionally, a building block is provided, specifically for:

[0160] Normalize the score map to obtain the initial pseudo mask;

[0161] Obtain the mask score of the initial pseudo mask, and use the mask score as the confidence score of the initial pseudo mask;

[0162] A non-maximum suppression operation is performed on the initial pseudo mask based on the confidence score to obtain a pseudo mask of the tail extension sample corresponding to the score mapping.

[0163] Optionally, the true value label includes a true mask and a true category label; the computing module is specifically configured to:

[0164] Inputting the head data sample, the tail data sample, and the tail extension sample into a preset instance segmentation network to obtain the predicted mask and predicted semantic category corresponding to the head data sample, the tail data sample, and the tail extension sample;

[0165] Based on the real mask, pseudo mask and predicted mask, the segmentation mask loss is obtained;

[0166] Based on the true category label, pseudo category label and predicted semantic category, the category loss is obtained.

[0167] Optionally, the optimization module is specifically configured to:

[0168] When the tail contrast loss, segmentation mask loss, and category loss meet the preset loss requirements, the preset instance segmentation network is used as the optimized instance segmentation network;

[0169] Otherwise, the network parameters in the preset instance segmentation network are adjusted to obtain an adjusted instance segmentation network;

[0170] The head data samples, tail data samples and tail extension samples are input into the adjusted instance segmentation network until the corresponding tail contrast loss, segmentation mask loss and category loss meet the preset loss requirements to obtain the optimized instance segmentation network.

[0171] It should be noted that the instance segmentation network optimization system provided in the above embodiment and the instance segmentation network optimization method provided in the above embodiment belong to the same concept, wherein the specific manner in which each module and unit performs operations has been described in detail in the method embodiment and will not be repeated here. In actual applications, the instance segmentation network optimization system provided in the above embodiment can allocate the above functions to different functional modules as needed, that is, divide the internal structure of the system into different functional modules to complete all or part of the functions described above, and this is not limited here.

[0172] A computing device according to an embodiment of the present application includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, some or all steps of the above-mentioned instance segmentation network optimization method are implemented.

[0173] Among them, the computing device can be a computer, and correspondingly, its program is computer software. The above-mentioned parameters and steps in a computing device of the present application can refer to the parameters and steps in the embodiment of the example segmentation network optimization method above, and will not be repeated here.

[0174] In an embodiment of the present application, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium. When the instructions are executed, the steps of the above-mentioned instance segmentation network optimization method are executed.

[0175] The computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0176] The technical solutions of the embodiments of the present disclosure can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes one or more instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method of the embodiments of the present disclosure. The aforementioned computer readable storage medium can be a non-transitory computer readable storage medium, including: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes, and can also be a transitory computer readable storage medium.

[0177] The flowcharts and block diagrams in the drawings illustrate the possible implementation architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In the flowcharts or block diagrams, each block can represent a module, a program segment or a part of code, and the module, the program segment or the part of code include one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from those noted in the drawings. For example, two blocks represented in succession can actually be executed in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams or flowcharts, and the combination of blocks in the block diagrams or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0178] Those skilled in the art know that the present application can be implemented as a system, a method or a computer program product. Therefore, the present disclosure can be embodied in the form of a complete hardware, a complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, which is generally referred to as "module" or "system" herein. In addition, in some embodiments, the present application can also be implemented in the form of a computer program product in one or more computer readable media, which includes computer readable program codes. The computer readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof.

[0179] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0180] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for optimizing an instance segmentation network, characterized in that: include: Acquire batches of tail data samples, and perform data augmentation on the tail data samples to obtain tail extended samples; Performing comparative learning on the tail data sample and the tail extension sample to obtain a tail comparative loss; Constructing a pseudo semantic label for the tail extension sample specifically includes: performing feature extraction on the tail extension sample to obtain dense features; downsampling the dense features to obtain reduced dimensionality features, and using the reduced dimensionality features as pseudo category labels for the tail extension sample; calculating a second cosine similarity between each of the reduced dimensionality features and each of the dense features; obtaining a score mapping for the tail extension sample based on the second cosine similarity and the dense features; calculating a pseudo mask for the corresponding tail extension sample based on the score mapping; and forming a pseudo semantic label based on the pseudo mask and the pseudo category label; Obtaining a batch of header data samples, wherein the header data samples and the tail data samples both contain true value labels, and the true value labels include true masks and true category labels; Inputting the head data sample, the tail data sample, and the tail extension sample into a preset instance segmentation network, and calculating the segmentation mask loss and the category loss based on the true value label and the pseudo semantic label, specifically comprising: inputting the head data sample, the tail data sample, and the tail extension sample into a preset instance segmentation network, obtaining the predicted mask and the predicted semantic category corresponding to the head data sample, the tail data sample, and the tail extension sample; obtaining the segmentation mask loss based on the true mask, the pseudo mask, and the predicted mask; and obtaining the category loss based on the true category label, the pseudo category label, and the predicted semantic category; Based on the tail contrast loss, the segmentation mask loss and the category loss, the preset instance segmentation network is adjusted to obtain an optimized instance segmentation network.

2. The method according to claim 1, characterized in that The performing comparative learning on the tail data sample and the tail extension sample to obtain the tail contrast loss includes: Performing feature extraction on the tail data sample to obtain tail data features; Performing feature extraction on the tail extension sample to obtain tail extension data features; Calculating a first cosine similarity between each tail extension data feature and each tail data feature to obtain a similarity set corresponding to each tail extension data feature; Obtaining a feature contrast loss of the tail extended data feature based on the tail extended data feature and the similarity set; Based on the tail extended data feature and the similarity set, performing feature enhancement on the tail extended data feature, and determining an enhancement contrast loss of the tail extended data feature; A tail contrast loss is formed based on the feature contrast loss and the enhancement contrast loss.

3. The method according to claim 2, characterized in that The obtaining, based on the tail extended data feature and the similarity set, a feature contrast loss of the tail extended data feature, includes: Taking the plurality of similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the plurality of similarity sets except the to-be-compared sets; The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples; The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples; A feature comparison loss is obtained based on tail extension data features corresponding to the positive sample, the negative sample, and the set to be compared.

4. The method according to claim 2, characterized in that The step of performing feature enhancement on the tail extended data feature based on the tail extended data feature and the similarity set, and determining the enhancement contrast loss of the tail extended data feature, includes: Taking the plurality of similarity sets as to-be-compared sets respectively, and obtaining first target similarities between the to-be-compared sets and other similarity sets in the plurality of similarity sets except the to-be-compared sets; The tail extended data features corresponding to other similarity sets whose first target similarity is greater than or equal to the preset similarity are taken as positive samples; The tail extended data features corresponding to other similarity sets whose first target similarity is less than the preset similarity are used as negative samples; Taking the first cosine similarity greater than a threshold in the set to be compared as the second target similarity; performing a feature aggregation operation on the tail data features corresponding to the second target similarity and the tail extended data features corresponding to the set to be compared to obtain enhanced features; An enhanced contrast loss of the tail extended data feature corresponding to the to-be-compared set is obtained based on the positive sample, the negative sample and the enhanced feature.

5. The method according to claim 1, wherein The calculating, based on the score mapping, a pseudo mask of the corresponding tail extension sample comprises: Normalizing the score map to obtain an initial pseudo mask; Obtaining a mask score of the initial pseudo mask, and using the mask score as a confidence score of the initial pseudo mask; A non-maximum suppression operation is performed on the initial pseudo mask based on the confidence score to obtain a pseudo mask of the tail extension sample corresponding to the score mapping.

6. The method according to any one of claims 1 to 5, characterized in that The step of adjusting the preset instance segmentation network based on the tail contrast loss, the segmentation mask loss, and the category loss to obtain an optimized instance segmentation network includes: When the tail contrast loss, the segmentation mask loss, and the category loss meet preset loss requirements, using the preset instance segmentation network as the optimized instance segmentation network; Otherwise, adjusting network parameters in the preset instance segmentation network to obtain an adjusted instance segmentation network; The head data sample, the tail data sample, and the tail extension sample are input into the adjusted instance segmentation network until the corresponding tail contrast loss, the segmentation mask loss, and the category loss meet preset loss requirements, thereby obtaining an optimized instance segmentation network.

7. A computing device comprising a memory, a processor, and a program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps of the instance segmentation network optimization method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the steps of an instance segmentation network optimization method according to any one of claims 1 to 6.

Images