Lightweight MotionBERT-based human body behavior recognition method

Abstract

The invention discloses a human behavior recognition method based on lightweight MotionBERT, and the method comprises the following steps: collecting a video data set required by a human behavior recognition task, namely video clips of different behaviors; preprocessing the data set; encoding the extracted skeleton sequence; in the MotionBERT model, a random mask and noise are applied to destroy a 2D skeleton, specific factors meeting preset conditions are removed, a double-flow network architecture is adjusted to construct an improved MotionBERT lightweight model, and a skeleton sequence is used as input; the preprocessed skeleton sequence is input into the improved MotionBERT lightweight model for training; a loss function FocalLoss is used to optimize the model; evaluating the performance of the trained improved MotionBERT lightweight model on the test set; and deploying the trained MotionBERT lightweight model to an embedded system, and carrying out human body behavior recognition. According to the method, on the premise of ensuring the accuracy of the operation result, the model is effectively lightened, and the operation resources are released.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and relates to a human behavior recognition method based on lightweight MotionBERT. Background Art

[0002] Human behavior recognition algorithms usually use computer vision technology and artificial intelligence algorithms to identify and classify the actions and postures of the detected objects. In recent years, the rapid development of computer technology and sensor technology has promoted the widespread application of human behavior recognition in the fields of intelligent monitoring and health management. At present, human behavior recognition algorithms usually use traditional artificial intelligence algorithms such as convolutional neural networks and recurrent neural networks to learn and classify human behavior data end-to-end. The MotionBERT model is a variant of the BERT model specifically used to process video data. It converts video data into frame-level feature sequences and inputs them into the BERT model to achieve effective encoding and semantic understanding of video data. The MotionBERT model can learn the motion information and contextual relationships in the video, thereby realizing the recognition of video actions. The architecture of the model includes key components such as video data input processing, time series modeling, and cross-modal learning. It encodes the video frame sequence through a multi-layer Transformer module and uses cross-modal learning technology to fuse text and video information, thereby improving the performance of the model in action recognition tasks and promoting its broad application prospects in the field of video action recognition. Therefore, it can be applied to video content understanding, video behavior recognition, video recommendation and other aspects, bringing new breakthroughs to video understanding tasks. Compared with traditional artificial intelligence algorithms such as convolutional neural networks and recurrent neural networks, the MotionBERT model has better modeling capabilities, stronger generalization capabilities and higher computational efficiency in human behavior recognition tasks, so it has great advantages in practical applications. However, the MotionBERT model usually requires a lot of computing resources to process video data, especially when the input sequence is long, the computational complexity of the model will increase further. This leads to difficulties in deployment in resource-constrained environments, such as mobile devices or embedded systems. In order to deploy the MotionBERT model on mobile devices and embedded systems to realize human behavior recognition, the algorithm model needs to be lightweight to make it more suitable for deployment and operation in resource-constrained environments.

[0003] Therefore, in view of the technical defects of the prior art, it is necessary to propose a solution to solve the technical problems of the prior art. Summary of the invention

[0004] The present invention is used to provide a human behavior recognition method based on lightweight MotionBERT, which reduces the complexity of the algorithm model by improving the GAU module and the hybrid approximate attention module instead of the MSA, releases computing resources, and is compatible with resource-constrained application scenarios.

[0005] To achieve the above object, the technical solution of the present invention is a human behavior recognition method based on lightweight MotionBERT, comprising the following steps:

[0006] S10, collects the video dataset required for the human behavior recognition task, i.e., video clips of different behaviors;

[0007] S20, preprocessing the data set;

[0008] S30, encoding the extracted skeleton sequence;

[0009] S40, in the MotionBERT model, the 2D skeleton is destroyed by applying random masks and noise, specific factors that meet the preset conditions are removed, and its two-stream network architecture is adjusted to build an improved MotionBERT lightweight model, taking the skeleton sequence as input;

[0010] S50, inputting the preprocessed skeleton sequence into the improved MotionBERT lightweight model for training;

[0011] S60, use the loss function FocalLoss to optimize the model;

[0012] S70, evaluate the performance of the trained improved MotionBERT lightweight model on the test set;

[0013] S80 deploys the trained MotionBERT lightweight model to the embedded system for human behavior recognition.

[0014] Preferably, the data set is preprocessed in S20, including format conversion and / or sampling and / or interception and / or video frame extraction.

[0015] Preferably, the video frame extraction specifically includes using a human posture detector to extract skeleton sequences from video frames for input into a human behavior recognition model.

[0016] Preferably, the S20 specifically includes the following steps:

[0017] S21, uses yolopose as an end-to-end human skeleton extractor;

[0018] S22, post-processing the results obtained by yolopose prediction.

[0019] Preferably, the S21 comprises the following steps:

[0020] S211, converting the video data into a bgr format;

[0021] S212, re-transform to a uniform resolution as the current video frame, and input the video frame into the yolopose model after coordinate transformation;

[0022] S213, through the deep convolutional network with CSPNet as the backbone, the target box, category and 17 skeleton key points are predicted. The prediction results include the location information and confidence of each point.

[0023] Preferably, the S22 specifically includes the following steps:

[0024] S221, filter out the confidence points less than 0.5, and filter out the skeleton key points outside the target frame;

[0025] S222, when the video is input, the coordinates are transformed, and each coordinate needs to be mapped back to the original coordinate system;

[0026] S223, through the NMS algorithm, when the overlap degree of the current two target frames is greater than 0.5, the frame with higher confidence is retained, and the frame with the largest confidence is selected from the remaining frames as the target frame of the current human body and the corresponding skeleton key points are attached.

[0027] Preferably, the S40 comprises the following steps:

[0028] S41, uses the MotionBERT model as the baseline model and retains its two-stream network architecture;

[0029] S42, the focused gated attention unit module F-GAU fuses the focused linear attention FLAT and the gated attention unit GAU to obtain the F-GAU module;

[0030] S43, the hybrid approximate attention module MMAT merges spatiotemporal features through feature fusion, and approximates the attention results through the linear gated unit GLU layer;

[0031] S44, adding information interaction between channels to improve the performance of the network, the formula is:

[0032] MAAT(m)=FC{βm+δST-Conv(GELU(FC(m)))} (2)

[0033] Where m is the result of the previous layer attention calculation, FC is the linear layer, β and δ are two variable parameters used to adjust the weight between local features and overall features, GELU is the activation function, and ST-conv() is a function used to calculate the spatiotemporal convolution and polynomial multiplication of the vector.

[0034] Preferably, the dual-stream network architecture in S41 includes a focus gated attention unit module F-GAU, a hybrid approximate attention module and an adaptive average pooling technology.

[0035] Preferably, the calculation formula of the F-GAU module in S42 is:

[0036]

[0037] In the formula, Q, K represent the queries of the input sequence, keys, s, n represent the length of the skeleton sequence, and relu 2 () is an activation function that replaces the traditional softmax operation. Represents the index, and takes the Q and K matrices bit by bit The power makes the attention matrix distribution difference more obvious. ε and σ are the feature norms of the Q and K matrices, which keep the feature norm unchanged and only change the direction of the feature, ultimately allowing the attention map to focus on more important features.

[0038] Preferably, the S60 includes: using the FocalLoss function as the loss function during the training process, and the calculation formula of the FocalLoss function is:

[0039] FL(p t )=-α(1-p t ) γ log(p t ) (3)

[0040] In the formula, p t is the predicted probability of the sample, α and γ are scaling factor coefficients.

[0041] The present invention has at least the following specific beneficial effects: In order to reduce the complexity of the behavior recognition model in the video, this paper proposes a lightweight human behavior recognition algorithm that improves MotionBERT. First, the original GAU module is improved to make the foreground and background more distinct, while maintaining the model performance and reducing the computational complexity. Secondly, a hybrid approximate attention module composed of a linear layer and a spatiotemporal convolutional layer is used to replace the traditional MSA in the deep layer. This approximation method not only effectively reduces the number of parameters of the model, integrates local and overall features, but also introduces information interaction between channels, further optimizing the expressive power of the model. Finally, the loss function is replaced with Focal Loss, which can pay more attention to samples that are difficult to classify during training, thereby improving the model's learning ability for key samples. Thereby, computing resources are released, making it compatible with resource-constrained application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The network structure of an improved MotionBERT of a human behavior recognition method based on lightweight MotionBERT according to an embodiment of the present invention;

[0043] Figure 2 A hybrid approximate attention module for a human behavior recognition method based on lightweight MotionBERT according to an embodiment of the present invention

[0044] Figure 3 This is the verification result of the focal loss of the human action recognition method based on lightweight MotionBERT in an embodiment of the present invention on a data set. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0046] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0047] Human behavior recognition is crucial for intelligent monitoring and health management, and MotionBERT can be effectively used to process video-related tasks. However, the MotionBERT algorithm model is large and usually requires a lot of computing resources to process video data. Especially when the input sequence is long, the computational complexity of the model will increase further. This leads to difficulties in deployment in resource-constrained environments, such as on mobile devices or embedded systems. In order to realistically deploy the MotionBERT model on mobile devices and embedded systems to realize human behavior recognition, the algorithm model needs to be lightweight to make it more suitable for deployment and operation in resource-constrained environments. In order to overcome the shortcomings of the prior art, the present invention considers proposing an improved network structure of MotionBERT, and its structure can be found in Figure 1 This model adopts the MotionBERT model as the baseline model and retains its two-stream network architecture.

[0048] In order to reduce the computational complexity of attention operations, the Gated Attention Unit (GAU) module is introduced. The Gated Attention Unit GAU module first uses relu 2 The () activation function replaces the traditional softmax operation. The execution of softmax often has a high computational cost, while the relu function can greatly reduce the computational overhead. Secondly, GAU integrates the self-attention operation with the feedforward network (FFN), making the calculation more efficient. In order to improve the performance of the gated attention unit GAU module in the spatiotemporal dimension, the focused linear attention FLAT and the gated attention unit GAU are fused to obtain the F-GAU module. The FLAT module can use the focusing function to make the feature distribution sharper, and then use the depthwise separable convolution to restore the rank of the attention matrix, so that the F-GAU module can make the result of the linear self-attention operation more accurate while ensuring the inference speed. The focused gated attention unit module F-GAU is used to optimize the spatiotemporal attention mechanism to enhance the flexibility of the model so that it can adapt to spatiotemporal data of different types and complexities. The alternating use of spatiotemporal F-GAU (SF-GAU and TF-GAU) constitutes the spatiotemporal feature extraction module.

[0049] In order to further reduce the redundancy of the attention model and reduce the amount of computation, the Mixed Approximation Attention module (MAAT) is used to improve the efficiency and performance of the attention mechanism. Its module structure is as follows: Figure 2As shown in the figure, this module designs spatial hybrid approximate attention S-MAAT and temporal hybrid approximate attention T-MAAT for the temporal and spatial features of long-term skeleton sequences. It merges spatiotemporal features through feature fusion, and adds residual connections to fuse local and global features, so that it can better identify action behaviors. After the hybrid approximate attention module MAAT module, a simple linear gated unit GLU layer is used to approximate the attention results. Finally, information interaction between channels is added to improve the performance of the network.

[0050] In order to reduce the redundancy of the classification head, the adaptive average pooling technique is used to further reduce the number of parameters and computational burden of the model.

[0051] In order to effectively solve the class imbalance problem of target detection, the FocalLoss function is used as the loss function. By introducing an adjustable scaling factor, the weights of samples that are easy to distinguish can be adaptively reduced during the training process, thereby quickly focusing the model's attention on samples that are more difficult to distinguish, reducing the impact of the class imbalance problem on model training and improving the performance of the target detection model.

[0052] In order to overcome the defects of the prior art, the present invention proposes a human behavior recognition algorithm based on lightweight MotionBERT. Since the traditional MotionBERT model has too large calculation and parameter volume, the present invention considers proposing an improved MotionBERT network structure, see Figure 1 The figure shows the improved network structure of MotionBERT. The model uses the MotionBERT model as the baseline model and retains its two-stream network architecture. It uses the focused gated attention unit F-GAU module to optimize the spatiotemporal attention mechanism and uses the hybrid approximate attention module MMAT to replace the attention calculation to reduce the number of parameters and computational complexity in the deep network. At the same time, by using the adaptive average pooling technology: Adaptive Average Pooling (AAP) + Feature Selection (FC) + FC, the redundancy of the classification head is successfully reduced, further reducing the number of parameters and computational burden of the model. The improved network structure of MotionBERT effectively improves the training and reasoning speed of the model while slightly reducing the classification accuracy, and greatly reduces the number of parameters and computational complexity of the model. Specifically, it includes the following steps:

[0053] S10, collects the video dataset required for the human behavior recognition task, i.e., video clips of different behaviors;

[0054] S20, preprocessing the data set, i.e., format conversion, sampling or interception, video frame extraction, etc. The video frame extraction process uses a human posture detector to extract skeleton sequences from video frames for input into a human behavior recognition model;

[0055] S30, encoding the extracted skeleton sequence;

[0056] S40, in the MotionBERT model, random masks and noise are applied to destroy the 2D skeleton and remove some specific factors, thereby improving the generalization of the model, and adjusting its two-stream network architecture to build an improved MotionBERT lightweight model, taking the skeleton sequence as input;

[0057] S50, inputting the preprocessed skeleton sequence into the improved MotionBERT lightweight model for training;

[0058] S60, optimize the model using an appropriate loss function FocalLoss;

[0059] S70, evaluate the performance of the trained improved MotionBERT lightweight model on the test set;

[0060] S80 deploys the trained MotionBERT model to the embedded system to realize human behavior recognition.

[0061] S20 includes the following steps:

[0062] S21 uses yolopose as an end-to-end human skeleton extractor. First, the video data must be converted into bgr format, and then reconverted to a uniform resolution (such as 640×640) as the current video frame. After the coordinate transformation, the video frame is input into the yolopose model, and the target box, category, and 17 skeleton key points are predicted through a deep convolutional network with CSPNet (Cross Stage Partial Network) as the backbone. The prediction results include the location information and confidence of each point.

[0063] S22, post-process the results predicted by yolopose. First, filter out those with a confidence level less than 0.5, and filter out the skeleton key points outside the target frame. Since the coordinates are transformed during video input, it is necessary to map each coordinate back to the original coordinate system. Finally, through the NMS (Non-maximum suppression) algorithm, when the overlap between the two current target frames is greater than 0.5, retain the frame with a higher confidence level, and select the frame with the highest confidence level from the remaining frames as the target frame of the current human body and attach the corresponding skeleton key points.

[0064] Wherein, step S40 includes the following steps:

[0065] S41 uses the MotionBERT model as the baseline model and retains its dual-stream network architecture. The structure includes the focus gated attention unit module F-GAU, the hybrid approximate attention module MMAT and the adaptive average pooling technology AAP+FC+FC.

[0066] S42, the focused gated attention unit module F-GAU fuses the focused linear attention FLAT and the gated attention unit GAU to obtain the F-GAU module. The calculation formula of this module is as follows:

[0067]

[0068] In the formula, Q, K represent the queries of the input sequence, keys, s, n represent the length of the skeleton sequence, and relu 2 () is an activation function that replaces the traditional softmax operation. Represents the index, and takes the Q and K matrices bit by bit The power makes the attention matrix distribution difference more obvious. ε and σ are the feature norms of the Q and K matrices, which keep the feature norm unchanged and only change the direction of the feature, ultimately allowing the attention map to focus on more important features.

[0069] S43, the hybrid approximate attention module MMAT merges spatiotemporal features (spatial hybrid approximate attention S-MAAT and temporal hybrid approximate attention T-MAAT) through feature fusion, approximates the attention results through a simple linear gating unit GLU layer, and finally adds information interaction between channels to improve the performance of the network. The calculation formula of this module is as follows:

[0070] MAAT(m)=FC{βm+δST-Conv(GELU(FC(m)))} (2)

[0071] In the formula, m is the result of the attention calculation of the previous layer, FC is the linear layer, β and δ are two variable parameters used to adjust the weight between local features and overall features. GELU is the activation function, and ST-conv() is a function used to calculate the spatiotemporal convolution and polynomial multiplication of the vector.

[0072] Wherein, step S60 includes the following steps:

[0073] S61, the FocalLoss function is used as the loss function during training, which adaptively reduces the weights of samples that are easy to distinguish, thereby quickly focusing the model's attention on samples that are more difficult to distinguish. The calculation formula of the FocalLoss function is as follows:

[0074] FL(p t )=-α(1-p t ) γ log(p t ) (3)

[0075] In the formula, p t is the predicted probability of the sample, α and γ are scaling factor coefficients.

[0076] Figure 2 This is the structure diagram of the hybrid approximate attention module MAAT, in which MHSA l-1 It is the result of the calculation of the previous two-stream network, GELU is the activation function (GELU(x) = x·Φ(x), Φ(x) is the Gaussian cumulative distribution function), S-Conv and T-Conv are spatial convolution and temporal convolution respectively.

[0077] Figure 3 The top1 classification accuracy (top1 ACC) using the focal loss function FocalLoss and the cross entropy loss function CELoss, as well as the validation loss (loss) using FocalLoss and CELoss. Direct use of FocalLoss sometimes causes misclassification and leads to unstable training. The experiment first uses CELoss for 25 rounds of training and then uses focal loss. Figure 3 It can be seen that using focal loss after a certain number of training rounds can further improve the accuracy of action classification and reduce function loss. This result shows that FocalLoss can focus on difficult samples and focus on learning, thereby obtaining more accurate results.

[0078] Table 1 shows the changes in the computational complexity and parameter count of the modified MotionBERT network structure. The results show that the training and reasoning speed of the modified model is slightly improved by replacing the multi-head self-attention module MSA (Multi-head Self-Attention Module, MSA) with the F-GAU module under the same hyperparameters. The training and reasoning speed of the modified model is 19.4% faster than that of the F-GAU, and the number of parameters is reduced by 9.8%. Compared with the baseline model, the training and reasoning speed is increased by 38.9%, and the number of parameters is reduced by 15.4%.

[0079] Table 1 Changes in calculation amount and parameter amount after modification

[0080]

[0081] Table 2 is a comparison of various model structure designs. The modified model has a similar computational complexity to the MSA model with a depth of 3 layers, and its accuracy is 0.8% higher than that of the MSA model. It has a 0.7% improvement in accuracy compared to the F-GAU model with the same depth of 5 layers, and the computational complexity is lower. Compared with MSA and the model with only S-GAU added, the accuracy and inference speed have improved. Combining Tables 1 and 2, it can be seen that the modified model not only has a reduced amount of computation and parameters, but also has improved recognition accuracy. Table 3 shows the impact of different loss functions on the classification results. After FocalLoss, the classification accuracy on the NTU-60Xsub and Xview datasets without fine-tuning is improved by 0.4%, and after fine-tuning, it is also improved by 0.2%.

[0082] Table 2 Changes in calculation amount and parameter amount after modification

[0083]

[0084] Table 3 shows the test results of the modified model on the NTU-60 dataset. The MotionBert model has a higher accuracy rate than other models. The proposed human action recognition algorithm based on lightweight MotionBERT has a slightly lower accuracy rate due to the simplified model structure, but it is better than other algorithms.

[0085] Table 3 Test results of different methods on NTU-60 (%)

[0086]

[0087] Table 4 shows the comparison of the number of parameters and the amount of computation of the modified model for the NTU-60 dataset. Under the premise that the time series length T is consistent, the number of parameters and FLOPs of the proposed model are significantly reduced.

[0088] Table 4 Comparison of parameter quantities and calculation quantities

[0089]

[0090] It can be seen from the above experimental results that compared with the traditional human behavior recognition algorithm, the optimization algorithm proposed in the present invention has a slightly lower classification accuracy, while greatly reducing the number of model parameters and the amount of calculation, so that it can run on mobile devices and embedded systems.

[0091] The above embodiments are only used to help understand the method and core idea of ​​the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0092] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A human behavior recognition method based on lightweight MotionBERT, characterized in that: The following steps are involved: S10, collects the video dataset required for the human behavior recognition task, i.e., video clips of different behaviors; S20, preprocessing the data set; S30, encoding the extracted skeleton sequence; S40, in the MotionBERT model, the 2D skeleton is destroyed by applying random masks and noise, specific factors that meet the preset conditions are removed, and its two-stream network architecture is adjusted to build an improved MotionBERT lightweight model, taking the skeleton sequence as input; S50, inputting the preprocessed skeleton sequence into the improved MotionBERT lightweight model for training; S60, use the loss function FocalLoss to optimize the model; S70, evaluate the performance of the trained improved MotionBERT lightweight model on the test set; S80 deploys the trained MotionBERT lightweight model to the embedded system for human behavior recognition.

2. The method according to claim 1, characterized in that In S20, the data set is preprocessed, including format conversion and / or sampling and / or interception and / or video frame extraction.

3. The method according to claim 2, characterized in that The video frame extraction specifically includes using a human posture detector to extract skeleton sequences from video frames for input into a human behavior recognition model.

4. The method according to claim 1, characterized in that: The S20 specifically includes the following steps: S21, uses yolopose as an end-to-end human skeleton extractor; S22, post-processing the results obtained by yolopose prediction.

5. The method according to claim 4, characterized in that The S21 comprises the following steps: S211, converting the video data into a bgr format; S212, re-transform to a uniform resolution as the current video frame, and input the video frame into the yolopose model after coordinate transformation; S213, through the deep convolutional network with CSPNet as the backbone, the target box, category and 17 skeleton key points are predicted. The prediction results include the location information and confidence of each point.

6. The method according to claim 4, characterized in that The S22 specifically includes the following steps: S221, filter out the confidence points less than 0.5, and filter out the skeleton key points outside the target frame; S222, when the video is input, the coordinates are transformed, and each coordinate needs to be mapped back to the original coordinate system; S223, through the NMS algorithm, when the overlap degree of the current two target frames is greater than 0.5, the frame with higher confidence is retained, and the frame with the largest confidence is selected from the remaining frames as the target frame of the current human body and the corresponding skeleton key points are attached.

7. The method according to claim 1, characterized in that The S40 comprises the following steps: S41, uses the MotionBERT model as the baseline model and retains its two-stream network architecture; S42, the focused gated attention unit module F-GAU fuses the focused linear attention FLAT and the gated attention unit GAU to obtain the F-GAU module; S43, the hybrid approximate attention module MMAT merges spatiotemporal features through feature fusion, and approximates the attention results through the linear gated unit GLU layer; S44, adding information interaction between channels to improve the performance of the network, the formula is: MAAT(m)=FC{βm+δST-Conv(GELU(FC(m)))} (2) Where m is the result of the previous layer attention calculation, FC is the linear layer, β and δ are two variable parameters used to adjust the weight between local features and overall features, GELU is the activation function, and ST-conv() is a function used to calculate the spatiotemporal convolution and polynomial multiplication of the vector.

8. The method according to claim 7, characterized in that The dual-stream network architecture in S41 includes a focused gated attention unit module F-GAU, a hybrid approximate attention module and an adaptive average pooling technology.

9. The method according to claim 7, characterized in that: The calculation formula of the F-GAU module in S42 is: In the formula, Q, K represent the queries of the input sequence, keys, s, n represent the length of the skeleton sequence, and relu 2 () is an activation function that replaces the traditional softmax operation. Represents the index, and takes the Q and K matrices bit by bit The power makes the attention matrix distribution difference more obvious. ε and σ are the feature norms of the Q and K matrices, which keep the feature norm unchanged and only change the direction of the feature, ultimately allowing the attention map to focus on more important features.

10. The method according to claim 1, characterized in that The S60 includes: using the FocalLoss function as the loss function during the training process, and the calculation formula of the FocalLoss function is: FL(p t )=-α(1-p t ) γ log(p t ) (3) In the formula, p t is the predicted probability of the sample, α and γ are scaling factor coefficients.

Images