A small sample under the condition of power line hidden danger target detection method

Abstract

A kind of power line hidden danger target detection method under small sample condition, by introducing double-path SKNet network, the main feature extraction network SKNet-ResNet152 is constructed in combination with ResNet152 network;The main feature extraction network and the region candidate network are combined to construct new Faster R-CNN++ network;Establish basic dataset and small sample dataset, use basic dataset to train network, obtain network parameter weight;Remove the last layer parameter, keep other layer parameters unchanged, using small sample dataset to fine-tune the last layer classifier and regressor parameter;Adopt cosine similarity to calculate classification result.The present application introduces double-path SKNet and cross-domain migration parameter adjustment mechanism, enhances the network feature extraction capability and the parameter reliability under the condition of sample imbalance, improves the self-adaptive ability of algorithm to small sample, multi-scale target, and is suitable for the target detection of power line hidden danger under small sample condition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power system inspection and disaster prevention, and relates to a power line hidden danger target identification method. BACKGROUND

[0002] The safe operation of power lines is a solid guarantee for the stable development of power grids. However, due to various reasons, power lines often have various safety hazards, such as insulator self-explosion, transformer oil leakage, bird nests, honeycombs, wire strand breakage, suspended foreign objects, heavy machinery construction such as excavators and cranes, etc., which are undoubtedly a threat to the safe operation of power grids.

[0003] With the rapid development of the economic society, the huge line and complex power grid environment lead to gradually increasing detection and maintenance costs. Relying on traditional detection means not only consumes manpower but also easily causes detection danger in the detection of high-voltage lines, and the detection accuracy of visual inspection is also easily reduced with the working time and complexity of the detection object, thereby causing an increase in detection costs and leading to power line safety accidents. Under this background, unmanned aerial vehicle power inspection can be carried out for 24 hours without interruption, has various forms, and has a low personnel casualty rate, thereby overcoming many shortcomings of manual power inspection and being popularized in more and more countries and regions to ensure the safe and stable operation of important power equipment and facilities such as power transmission lines with higher inspection efficiency. However, the picture data collected by the unmanned aerial vehicle is huge, and manual investigation will consume a lot of time, has low efficiency, is easy to cause misjudgment and omission, is difficult to accurately find existing safety hazards, causes delay in fault handling, and inevitably causes losses, so it is imperative to develop intelligent power line inspection technology based on artificial intelligence technology.

[0004] Image classification technology using deep learning technology has been widely applied, and target detection, target segmentation, anomaly detection, and target tracking in deep learning have also made great progress and are applied in various fields. Power line detection has various complex natural backgrounds, differences in imaging conditions, multi-scale detection objects, and different clarity of different objects in the same image under the same focal length, and deep learning technology can solve these problems. However, the commonly used deep learning technology depends on the learning of a large number of samples, and only when the sample quantity is large enough can the trained model meet the accuracy requirements. In real applications, there are often problems such as difficulty in obtaining samples and high labeling costs, for example, in power line inspection, some safety hazards or equipment faults cannot be obtained in large quantities (such as abnormal suspended objects on the line, wire breakage, bird nests, and large machinery construction), and some fault image data also need to be labeled by professional technical personnel (such as insulator self-explosion, shock absorber falling off / shift / displacement, etc.), which brings difficulties to the establishment of a sample library and cannot meet the training requirements of commonly used deep learning network models, so it is necessary to study small sample target detection.

[0005] Current research on few-sample object detection faces numerous challenges. A limited number of samples easily leads to a decline in the performance of feature extractors and classifiers. Traditional deep learning training methods do not conform to the real-world environment of few-sample object detection, are unsuitable for training few-sample object detection networks, and cannot yield few-sample object detection networks with excellent classification performance. Only by solving these problems can we obtain few-sample object detection networks with better classification results. Due to the small sample size and overly complex networks, directly applying deep learning algorithms to these situations can easily lead to overfitting, i.e., high training accuracy but low testing accuracy. Therefore, object detection under few-sample conditions is of great significance for the intelligent detection of safety hazards in power lines and is one of the urgent problems to be solved in the field of computer vision.

[0006] In conclusion, research on small-sample target detection can reduce the dependence of traditional deep learning networks on large amounts of training data, which is conducive to improving the accuracy of power line safety hazard detection, better ensuring the safe operation of power lines, and better meeting the needs of the country's future scientific and technological development. Summary of the Invention

[0007] This invention focuses on typical safety hazards during power line inspections and proposes a method for detecting power line hazards under small sample conditions based on attention mechanisms and cross-domain transfer learning. A small sample dataset is constructed to address safety hazards such as transformer oil leaks, bird nests, honeycomb structures, hanging foreign objects, and heavy machinery construction (e.g., bulldozers, cranes). Faster R-CNN based on the Res152 network is selected as the main network for target detection. The PASCAL VOC 2007 dataset is used as a benchmark dataset for training the main network. The last layer of the main network is fine-tuned, and other network parameters are frozen. Training continues on a dataset containing small sample categories, thus constructing a power line hazard identification model suitable for small sample conditions. Experimental results demonstrate that using the Faster R-CNN network based on transfer learning and fine-tuning to identify power line safety hazards under small sample conditions is feasible.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] The present invention discloses a method for detecting potential hazards in power lines under small sample conditions, comprising the following steps:

[0010] S1. Construct the dataset, including a benchmark dataset for pre-training and a small sample dataset of power line hazard targets for transfer learning.

[0011] S2. A dual-channel SKNet attention module is introduced to preprocess the image, quickly extract features, and generate a feature map. For example... Figure 1As shown, SKNet attempts to find the most suitable convolutional kernels of different sizes for different images, dynamically generating kernels for each image. Unlike the multi-scale approach in Inception, SKNet allows the network to choose the appropriate scale itself. Figure 1 As shown.

[0012] The specific process is as follows:

[0013] S2.1. Perform grouped convolution (operation) on the input image using two convolution kernels of different sizes. The use of convolution kernels of different sizes is mainly to improve accuracy.

[0014] S2.2 After the process in S21 is divided into two branches, the results of the two branches are then merged.

[0015] S2.3. Using two softmax normalization functions, the weight information of each channel is regressed from the result of S22. Then, the two weight matrices are weighted on both paths and summed to obtain the output vector, i.e., the feature map. Because SKNet uses different convolutional kernels, it can adaptively adjust its receptive field, thus achieving higher accuracy than SENet.

[0016] S3. Using ResNet152 as the base network, replace the convolutional modules with SKNet to improve the residual network structure, forming a new SKNet-ResNet152 network structure, such as... Figure 2 As shown.

[0017] The specific process is as follows:

[0018] S3.1 and Conv1 remain unchanged from the ResNet152 network;

[0019] In S3.2 and Conv2, the 3×3 convolutional module is replaced with SKNet, while everything else remains unchanged, forming a new SKNet-Conv2_x module;

[0020] In S3.3 and Conv3, the 3×3 convolutional module is replaced with SKNet, while everything else remains unchanged, forming a new SKNet-Conv3_x module;

[0021] In S3.4 and Conv4, the 3×3 convolutional module is replaced with SKNet, while everything else remains unchanged, forming a new SKNet-Conv4_x module;

[0022] In S3.5 and Conv5, the 3×3 convolutional module is replaced with SKNet, while everything else remains unchanged, forming a new SKNet-Conv5_x module;

[0023] S3.6, and finally, the average pooling, full connection, and normalization operations remain unchanged, forming a new SKNet-ResNet152 network structure.

[0024] S4. Based on the SKNet-ResNet152 network and region candidate network in S3, construct a Faster R-CNN++ network that incorporates SKNet, such as... Figure 3 As shown.

[0025] The specific process is as follows:

[0026] S4.1. Perform convolution operations on the input image data using the Conv1 module to generate features. Figure 1 ;

[0027] S4.2, Features Figure 1 Convolutional operations were performed using three sets of SKNet-Conv2_x modules to generate features. Figure 2 ;

[0028] S4.3, Features Figure 2 Features were generated using convolutional operations with 8 SKNet-Conv3_x modules. Figure 3 ;

[0029] S4.4, Features Figure 3 Convolutional operations were performed using 36 SKNet-Conv4_x modules to generate features. Figure 4 ;

[0030] S4.5, Features Figure 4 As input, the region candidate network is used to process and generate candidate region information;

[0031] S4.6, Features Figure 4 Region pooling is performed on candidate region information to generate features. Figure 5 ;

[0032] S4.7, Features Figure 5 Three sets of SKNet-Conv5_x modules are used for convolution operations, and the logistic regression function is used to obtain the final classification and the final object detection box position.

[0033] The region candidate network (GAN) preprocesses the image regions to be processed, reducing subsequent computation. It mainly consists of two operations:

[0034] a. Suggested region classification: Perform region segmentation on the feature map and distinguish between foreground and background regions;

[0035] b. Boundary Regression: Extracts the approximate coordinates of the foreground region. During region candidate network training, the loss values ​​from both parts need to be summed. Let p be the probability that the anchor box predicts the target.i p i * represents the label value of the foreground sample, 1 for positive samples and 0 for negative samples; N cls It is the size of a batch, N reg The total number of anchor frames, λ is the balance ratio of the two loss values, and the log loss between the foreground and background is L. cls (p i ,p i *), t i t is the parametric coordinate of the i-th predicted anchor box. i * represents the parameterized coordinates of the manually labeled anchor box corresponding to the i-th predicted anchor box, R is the smoothing L1 function, and the loss function of Faster R-CNN is shown in Equations 1, 2, and 3.

[0036]

[0037]

[0038]

[0039] S5. Calculate the network's loss value. The network's loss function consists of three parts: a. Region candidate network loss L rpn : Depends on whether foreground and background are distinguished; b. Classification branch loss L cls c. Using cross-entropy loss function; d. Boundary regression loss L loc : Use the smoothed L1 loss function.

[0040] In terms of training method, the feature extraction network is trained using only the basic dataset, and the joint training loss is:

[0041] L = L rpn +L cls +L loc (4)

[0042] L rpn Including binary classification loss for candidate region foreground and background, and first regression loss for region optimization, L cls It is the cross-entropy loss of the last classification bundle, L loc It is the smoothed L1 loss of the regression bundle.

[0043] S6. Cross-domain transfer learning and parameter fine-tuning. Using a benchmark dataset and a few-sample dataset, a few-sample object detection model is built using a two-stage training and fine-tuning mechanism, such as... Figure 4 As shown.

[0044] The specific process is as follows:

[0045] S6.1 Based on the above network structure, the network is trained using the basic dataset to obtain a preliminary network model;

[0046] S6.2. Based on the network model obtained in S61, keep the parameters of the entire feature extractor unchanged and remove the parameters of the last layer of the model.

[0047] S6.3. Fine-tune the parameters of the classifier and regressor in the last layer of the model using a small sample dataset to obtain the weights of the small sample classes, such as... Figure 5 As shown;

[0048] S6.4. Perform transfer learning training on a balanced dataset containing a benchmark dataset and a small sample dataset to achieve fine-tuning of parameters;

[0049] S6.5 Fine-tuning loss calculation, the loss function is the same as in S4.

[0050] S7. Finally, cosine similarity classification is used, with a weight matrix [w1, w2, ..., w...]. c ],w i This represents the weight vector for each category, and the output is the similarity score S. The formula for calculating the elements in S is:

[0051]

[0052] S i,j Let α represent the similarity score between the weight vectors of the i-th proposed region and the j-th class in the input image x, where α is a scaling factor, set to 10 here. Using cosine similarity classification with strength-level feature normalization can reduce intra-class variance and improve the detection accuracy of new classes.

[0053] This invention focuses on the scientific problems of feature extraction and target detection in complex images under small sample conditions. It aims to establish a target detection model for power line safety hazards under small sample conditions, thereby improving the accuracy and stability of the detection model. Based on a combination of attention mechanisms, Faster R-CNN, and meta-learning methods, this invention studies feature extraction strategies and parameter adjustment mechanisms for complex images under small sample conditions. It constructs a small sample target detection model based on attention mechanisms and cross-domain transfer learning, achieving accurate detection of power line safety hazards under such conditions. This method has significant theoretical and practical implications.

[0054] Compared with the prior art, the present invention has the following beneficial technical effects:

[0055] 1) This invention introduces dual-path SKNet in the feature extraction stage, which enhances the network's ability to detect targets at multiple scales and is more suitable for power line inspection images with complex backgrounds.

[0056] 2) In view of the high complexity of power line inspection image data, this invention combines two-stage target detection and two-stage parameter fine-tuning mechanism to enhance the network’s feature extraction capability for small sample images and improve the reliability of network parameters on small sample datasets.

[0057] 3) To address the issue of low detection accuracy in certain categories due to single samples or extremely imbalanced samples, this invention introduces cross-domain transfer learning and parameter adjustment strategies to further improve the detection accuracy and overall stability of the model under small / imbalanced sample conditions. Attached Figure Description

[0058] Figure 1 This is a diagram of the dual-path SKNet network structure.

[0059] Figure 2 The residual network structure diagram of SKNet is introduced for this invention.

[0060] Figure 3 The network structure diagram of SKNet Faster R-CNN+ResNet152 is introduced for this invention.

[0061] Figure 4 This is a schematic diagram of the two-stage training and parameter fine-tuning steps of the present invention.

[0062] Figure 5 This is the parameter adjustment step for cross-domain transfer learning in this invention. Detailed Implementation

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0064] A power line hazard detection method based on attention mechanism and cross-domain transfer learning under small sample conditions includes the following steps:

[0065] S1. Construct the dataset, using the PASCAL VOC 2007 dataset as the base dataset; in addition, create a small sample dataset using collected images of potential hazards in power lines, mainly including 6 categories: transformer oil leakage, hanging foreign objects, fire, bird nests, beehives, and large machinery, with 10 samples in each category.

[0066] S2. Construct a dual-channel SKNet attention network module, as shown in Table 1. The SKNet processing procedure is as follows:

[0067] 1) Perform convolution processing on the input data to obtain a preliminary feature map.

[0068] 2) Perform convolution operations on the feature maps after convolutional layer processing using 3*3 and 7*7 convolution kernels respectively to obtain outputs U1 and U2.

[0069] 3) The multi-scale features from the previous step are fused by adding corresponding elements of the feature maps, thus fusing the results of the two branches. U is a feature map of size C*H*W (C represents the number of channels, H represents the height, and W represents the width) that incorporates information from multiple receptive fields. Then, by averaging along the H and W dimensions, a vector s of size C*1*1 is obtained, representing the importance of each channel.

[0070] 4) Apply a fully connected layer to the C*1*1 vector to perform a linear transformation to obtain a Z*1*1 information z (as shown in equation (5). Then, use three linear transformations to restore the vector from Z dimensions to C dimensions and extract the information of each channel dimension.

[0071] z = F fc (s)=δ(B(W s (5)

[0072] Where: z∈R d×1 δ is the ReLU function; B is batch normalization, W∈R d×C , where d = max(C / r, L), r represents the reduction ratio, and L is the minimum value of d.

[0073] 4) Softmax normalization is used to obtain the corresponding score representing the importance of each channel. This score is then multiplied by the corresponding U1 and U2 to obtain A1 and A2. The two modules are then added together and fused to obtain V. V, relative to U, has undergone information refinement and incorporates information from multiple receptive fields. Let a and b be the two weight matrices of Select, A, B ∈ R. C×d A i Let a represent the i-th row of A. i B is the i-th element of a. i b i and A i Similarly, and a i +b i =1, and the final feature mapping V is as shown in equations (6) and (7) (where e is a natural number).

[0074] Vi = a i ×A1+b i ×A2 (6)

[0075]

[0076] Table 1 SKNet Network Structure

[0077]

[0078] Where M represents the number of network branches in step 2), G represents the number of groups of group convolutions in step 3), and r represents the reduction rate of the fully connected layer in step 3).

[0079] S3. Replace the convolutional modules in the ResNet152 network with SKNet to construct the SKNet-ResNet152 network module, as shown in Table 2. The specific process is as follows:

[0080] 1) Conv1 uses a 7×7 convolutional module with a stride of 2 and 64 channels;

[0081] 2) Conv2 first performs 3×3 max pooling with a stride of 2; then it replaces the 64-channel 3×3 convolutional module in the Conv2 module with SKNet to form a new SKNet-Conv2_x module.

[0082] 3) Replace the 128-channel 3×3 convolutional module in the Conv3 module with SKNet to form a new SKNet-Conv3_x module;

[0083] 4) Replace the 256-channel 3×3 convolutional module in the Conv4 module with SKNet to form a new SKNet-Conv4_x module;

[0084] 5) Replace the 512-channel 3×3 convolutional module in the Conv5 module with SKNet to form a new SKNet-Conv5_x module;

[0085] 6) Finally, average pooling, full connection, and normalization operations are performed to form a new SKNet-ResNet152 network structure with a total of 152 layers.

[0086] Table 2 SKNet-ResNet152 Network Structure

[0087]

[0088] S4. Based on the SKNet-ResNet152 network, construct a new Faster R-CNN++ network and use the new network to perform feature extraction processing on the image. The specific process is as follows:

[0089] 1) The input image data is processed using the Conv1 module to generate 112×112 features. Figure 1 ;

[0090] 2) Features Figure 1 The SKNet-Conv2_x modules were used for processing to generate 56×56 features. Figure 2 ;

[0091] 3) Features Figure 2The data was processed using 8 sets of SKNet-Conv3_x modules to generate 28×28 features. Figure 3 ;

[0092] 4) Features Figure 3 The data was processed using 36 SKNet-Conv4_x modules to generate 14×14 features. Figure 4 ;

[0093] 5) Features Figure 4 As input, the region candidate network is used to process and generate candidate region information;

[0094] 6) Features Figure 4 Perform region-of-interest pooling and merge candidate region information;

[0095] 7) The fused information is processed using three sets of SKNet-Conv5_x modules, and then processed using average pooling, fully connected, logistic regression functions, etc., to obtain the final classification and the final target detection box position.

[0096] S5. Calculate the joint loss value of the network.

[0097] S6. Network Training. The specific training steps are as follows:

[0098] 1) Initialize the parameters of each layer of the Faster R-CNN++ network using a Gaussian distribution with a mean of 0 and a variance of δ.

[0099] 2) Set the model hyperparameters.

[0100] Table 3 Model Hyperparameters

[0101]

[0102] 3) Training was performed using the base dataset. The base dataset contains 20 categories and a total of 5371 training images.

[0103] 4) In this example, 9 anchor points are set, corresponding to the size of 128×128, 256×256, and 512×512 pixels, and the aspect ratios of 1:1, 1:2, and 2:1, respectively. The Faster R-CNN++ network is approximately jointly trained to obtain the trained model on the basic dataset.

[0104] 5) Keep the parameters of the entire feature extractor in the model unchanged, and remove the parameters of the last layer of the model.

[0105] 6) Train the model using a small sample dataset through transfer learning to fine-tune the parameters of the classifier and regressor in the last layer, obtaining the weights of the small sample class. The training iterations are 300, with other training parameters remaining unchanged.

[0106] 7) Fine-tuning loss calculation, the loss function is the same as in 4).

[0107] 8) Cosine similarity calculation: By calculating the similarity score for each category, the output category with the highest similarity is obtained. The results are shown in Table 4:

[0108] Table 4 Training results for the small sample dataset

[0109]

[0110] In summary, this invention improves the Faster R-CNN network by introducing a dual-path SKNet, thereby enhancing the network's ability to detect targets at multiple scales. Furthermore, through cross-domain transfer learning and a two-stage parameter fine-tuning mechanism, it improves the network's ability to extract features from small sample data and the reliability of learning parameters, thus enabling the detection of safety hazards in power line inspections under small sample conditions.

[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A method for detecting power line hidden danger targets under small sample conditions, characterized in that The method comprises the following steps: S1, constructing a data set, including a benchmark data set for pre-training and a power line hidden danger target small sample data set for transfer learning; S2, introducing a double-channel SKNet attention module, pre-processing the image, quickly extracting features and generating a feature mapping graph: S2.1, using two convolution kernels of different sizes to perform grouped convolution on the input image; S2.2, after being processed by S2.1, the results of the two branches are fused; S2.3, using two softmax normalization functions, the results of S2.2 are regressed to obtain the weight information of each channel, then the two weight matrices are weighted processed, and the output vector, i.e., the feature mapping, is obtained by summation; S3, taking ResNet152 as the basic network, replacing the convolution module with SKNet to improve the residual network structure, forming a new SKNet-ResNet152 network structure: S3.1, Conv1 remains unchanged from ResNet152 network; S3.2, in Conv2, replace the 3×3 convolution module with SKNet, and other parts remain unchanged, forming a new SKNet-Conv2_x module; S3.3, in Conv3, replace the 3×3 convolution module with SKNet, and other parts remain unchanged, forming a new SKNet-Conv3_x module; S3.4, in Conv4, replace the 3×3 convolution module with SKNet, and other parts remain unchanged, forming a new SKNet-Conv4_x module; S3.5, in Conv5, replace the 3×3 convolution module with SKNet, and other parts remain unchanged, forming a new SKNet-Conv5_x module; S3.6, the last average pooling, full connection, normalization and other operations remain unchanged, forming a new SKNet-ResNet152 network structure; S4, based on the region candidate network and the SKNet-ResNet152 network in S3, a Faster R-CNN++ network with SKNet is constructed; the image is processed by the Faster R-CNN++ network for feature extraction, comprising the following steps: S4.1, the input image data is processed by the Conv1 module for convolution operation to generate feature map 1; S4.2, the feature map 1 is processed by 3 groups of SKNet-Conv2_x modules for convolution operation to generate feature map 2; S4.3, the feature map 2 is processed by 8 groups of SKNet-Conv3_x modules for convolution operation to generate feature map 3; S4.4, the feature map 3 is processed by 36 groups of SKNet-Conv4_x modules for convolution operation to generate feature map 4; S4.5, the feature map 4 is taken as input and processed by the region candidate network to generate candidate region information; S4.6, the feature map 4 and the candidate region information are subjected to region of interest pooling to generate feature map 5; S4.7, the feature map 5 is processed by 3 groups of SKNet-Conv5_x modules for convolution operation, and a logistic regression function is used to obtain the final classification and the final target detection frame position; S5, calculate the loss value of the network: the loss function of the network includes three parts: a. regional candidate network loss L rpn : depending on whether the foreground and background are distinguished; b. classification branch loss L cls : cross-entropy loss function is adopted; c. bounding box regression loss L loc : smooth L1 loss function is adopted; In terms of training method, the feature extraction network is trained using only the basic dataset, and the joint training loss is: （4） L rpn The first loss L1 is a binary classification loss including the foreground and background of the candidate region, and the second loss L2 is a regression loss for region optimization. cls The first loss L1 is a binary classification loss including the foreground and background of the candidate region, and the second loss L2 is a regression loss for region optimization. loc The first loss L1 is a binary classification loss including the foreground and background of the candidate region, and the second loss L2 is a regression loss for region optimization. S6. Cross-domain transfer learning and parameter fine-tuning: Using a benchmark dataset and a small sample dataset, a two-stage training and fine-tuning mechanism is employed to build a small sample object detection model. S6.1 Based on the above network structure, the network is trained using the basic dataset to obtain a preliminary network model; S6.

2. Based on the network model obtained in S6.1, keep the parameters of the entire feature extractor unchanged and remove the parameters of the last layer of the model. S6.

3. Fine-tune the parameters of the classifier and regressor in the last layer of the model using a small sample dataset to obtain the weights of the small sample classes; S6.

4. Perform transfer learning training on a balanced dataset containing a benchmark dataset and a small sample dataset to achieve fine-tuning of parameters; S6.5 Fine-tuning loss calculation, the loss function is the same as in S4; S7, finally, cosine similarity classification is adopted, and a weight matrix [w1, w2,..., w c ], w i represents a weight vector of each category, and the output is a similarity score S. The elements in S are calculated according to the following formula: （5） S i,j similarity score of the weight vector of the i-th proposal region of the input image x and the j-th class, and a is a scaling factor, which is set to 10; The region candidate network described in S4 includes two parts of operation: a. Suggested region classification: Perform region segmentation on the feature map and distinguish between foreground and background regions; b. Border regression: extract the position coordinates of the foreground region; the loss value of the two parts is added when the region candidate network is trained; the probability of setting the anchor frame prediction as the target is p i , p i * is the label value of the foreground, and the positive sample is 1 and the negative sample is 0; N cls is the size of a batch, N reg is the total number of anchor frames, λ is the balance ratio of the two loss values, and the logarithmic loss of the foreground and background is L cls (p i , p i *), t i is the parameterized coordinate of the i th predicted anchor frame, t i * is the parameterized coordinate of the i th predicted anchor frame corresponding to the artificial labeled anchor frame, and R is the smooth L1 function. The loss function of the Faster R-CNN is shown in formulas 1, 2 and 3: ， （1） ， （2） 。 （3）。

Images