A helmet wearing detection method based on implicit expression
By combining feature map pyramid network and path aggregation network, using implicit and explicit information to extract features, the problem of low detection accuracy of safety helmets on construction sites is solved, and the effect of high accuracy and rapid detection is achieved.
Patent Information
- Application Number
- CN202210062828.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-19
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-01-19
AI Technical Summary
The existing safety helmet wear detection system has low detection accuracy at the construction site and is greatly affected by environmental interference. The model training effect is poor during multi-harmet detection, making it difficult to achieve high accuracy and rapid detection in complex scenarios.
Using end-to-end neural network based on feature map pyramid network and path aggregation network, combining implicit and explicit information, extract local and contextual features, and feature alignment and prediction refinement through implicit information guidance module to improve detection accuracy and speed .
High-accurate safety helmet wear inspection is achieved in complex construction sites, with fast detection speed and can detect whether workers wear safety helmets correctly in real time, with an accuracy rate of 95.9%. It is suitable for inspections in multiple scenarios and different safety helmet sizes.
Smart Images

Figure CN114463676B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of helmet wearing detection systems, and in particular to a helmet wearing detection method based on implicit expression. Background Art
[0002] As we all know, control, automation, artificial intelligence and other fields have received widespread attention from the society in recent years. In any construction operation, the wearing of helmets by construction workers is the most important requirement and measure for production safety. Using manpower to monitor the real-time status of workers wearing helmets will greatly increase labor costs, and missed detections will also occur, because manual monitoring to check whether workers have illegal operations can easily cause fatigue. Due to problems such as perspective distortion, scale changes, internal occlusion and complex scenes at the construction site, the accuracy of safety helmet wearing detection is still poor. Therefore, the safety helmet wearing detection system in construction occasions has always been a very challenging problem.
[0003] Since helmets are small in size and in complex environments, with dim and obstructed light, traditional convolutional neural networks are difficult to process directly. Therefore, previous research work has detected helmets by using some details and features. For example, the histogram of oriented gradients (HOG) features and color histograms are used to detect pedestrians and the correct wearing of helmets. However, this method is easily affected by the environment, resulting in extremely poor detection accuracy. It also relies on sensors other than visual sensors, and its high cost and low robustness make it impossible to apply on a large scale. It was not until 2020 that Siebert et al. proposed the use of multi-scale feature pyramids to detect helmet wearing (FWSiebert, HHLin, Detecting motorcycle helmet use with deeplearning, Accident analysis and prevention, Vol. 134, 105319 (1)-105319 (11), 2020.). Its basic idea is to effectively extract features through convolutional neural networks, strengthen and utilize multi-scale features to obtain a more expressive feature map containing multi-scale helmet target area information, and finally use two convolutional layer networks with the same structure but different convolution kernel sizes on the feature map set of the feature map pyramid network to complete the target box category classification and prediction box position regression tasks. However, the number of available helmet data sets is small, the scene is single, and it cannot be well applied to actual complex construction sites, that is, the accuracy and detection speed still cannot meet the requirements. At the same time, since the sizes of helmets displayed in the image are not the same, the detection of multiple helmets, that is, multiple tasks, cannot train the model well, because when encountering the problem of multiple helmet detection, the output detection used for predicting close-range helmets (larger objects) and long-range helmets (smaller objects) together is also different. Therefore, during the training process, there is often a pull between the tasks of predicting different helmet sizes, resulting in poor training results. Moreover, in the training of multiple scenes and multiple different helmet sizes, Kernel space alignment is a difficult problem that often misleads the training direction and leads to poor training results.
[0004] Therefore, we propose an end-to-end neural network for helmet wearing detection algorithm that captures local fine-grained feature information while taking into account edge feature information. We propose a framework based on feature graph pyramid network and path aggregation network to combine implicit and explicit information to enhance the ability of contextual feature information. Using the proposed network, an implicit expression-based helmet wearing detection algorithm is proposed, which can significantly improve detection accuracy and operation time. While ensuring high accuracy, the detection speed can still be guaranteed, realizing real-time detection of whether workers are wearing helmets correctly in complex construction site scenes. Summary of the invention
[0005] The present invention mainly uses a deep learning method to detect workers wearing helmets to obtain a network with higher detection accuracy in a construction environment. First, an image of a worker wearing or not wearing a helmet is input. Through Mosaic data enhancement and adaptive image scaling, a series of preprocessing is performed on the input image. The local feature extraction module is used in the feature graph pyramid network to extract edge features, and the implicit information guidance module is used to align the feature graph pyramid network features. We add and multiply the feature pyramid map fused by the feature graph pyramid network and the implicit representation, which means that the kernel space is translated, rotated and scaled. Then, the gradient information is transmitted using different transmission paths through the context feature aggregation module, and the context features are aggregated to achieve global feature fusion. In the path aggregation network, all two feature maps of the same size are spliced in the channel dimension. After two path aggregation network structures, we resample the feature map to its original size to improve the helmet positioning capability. The feature graph pyramid network layer can capture strong semantic features from top to bottom, while the path aggregation network conveys strong positioning features from bottom to top. By combining these two modules, the function of helmet positioning can be well completed.
[0006] We use implicit information to guide the module applied to the output module, mainly for prediction refinement and learning of different helmet sizes. At the same time, the error term is modeled with the combined information to guide the network training process.
[0007] In order to achieve the above object, the present invention provides the following scheme:
[0008] A helmet wearing detection method based on implicit expression, the method comprising:
[0009] Step 1: Preprocess the input helmet image data;
[0010] Step 2: Construct a neural network for the target box category classification and prediction box position regression tasks that combines implicit and explicit information;
[0011] Step 3: Train the network model;
[0012] Step 4: Use the trained classification and regression task network models for evaluation.
[0013] The data preprocessing specifically includes the following steps:
[0014] Step 1: Dataset: Use an RGB camera to capture photos of workers wearing or not wearing helmets at the construction site and use annotation tool software such as Laelbox to annotate and create a dataset, or use existing public datasets to construct a training dataset.
[0015] Step 2: Preprocessing: Mosaic data enhancement is used to randomly splice and randomly scale the input data to achieve data amplification and enhancement. At the same time, the input image is scaled to the input size of the network using the adaptive image scaling method for normalization.
[0016] The neural network for constructing the target box category classification and prediction box position regression tasks combining implicit and explicit information specifically includes the following modules:
[0017] Module 1: Local feature extraction module
[0018] The local feature module is reflected in the extraction of local features across stages on the channel by integrating the gradient changes from beginning to end into the feature map. The specific operation is to split the feature map into two parts, one part uses the convolution operation with a convolution kernel of 1×1 and a step size of 1, and the other part inputs the feature map into two convolution operations with a convolution kernel of 1×1 and a step size of 1, a batch norm operation and a SILU activation function, and then uses the residual convolution structure to extract local features across stages, and then uses the convolution operation with a convolution kernel of 1×1 and a step size of 1 to adjust the number of channels, and then this part is spliced with the output result of the previous part in the channel dimension.
[0019] Module 2: Contextual Feature Aggregation Module
[0020] Context feature aggregation refers to the use of three filter sizes of 2×2, 4×4, and 6×6, with a step size of 1, for the input image. The pooling structure uses a parallel transmission path to transmit the gradient information, and the gradient flow information can be quite different, thereby optimizing the repeated gradient information. Finally, it is combined with the input feature map through a convolution operation with a convolution kernel of 1×1 and a step size of 1, and the feature aggregation map is output to enhance the feature representation capability. The calculation formula of the multi-scale context aggregation module is as follows:
[0021] Output = conv(concat(φ max (conv(f));φ max (conv(f));φ max (conv(f));conv(f))+conv(f))
[0022]
[0023] Where f represents the input feature map, conv(f) represents the normal convolution operation with a convolution kernel size of 1*1 and a stride of 1 on the input image f, and concat(x1; x2; ...; x n ) represents the input feature map {x1,x2,…,x n}Splicing along the channel direction, φ max The calculation is shown in the formula, C represents the total number of channels of the feature map, c represents the number of channels of the feature map, φ c represents the cth channel of the feature map, (h,w) represents the position in space, φ c (h,w) represents the feature information of the cth channel located at the (h,w) coordinate in the feature map.
[0024] Module 3: Implicit Information Guidance Module
[0025] The implicit information guidance module represents a set of constant tensors initialized before the training or prediction phase begins and added to the feature map pyramid network kernel space alignment, which is applied to the output module to predict the refinement and input image of the helmet size displayed as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, 1 / 2 to 1 multi-objective learning of the entire image.
[0026] Implicit information is added to the output module for refinement. The specific operation is to implicitly express the offset of the center coordinate predicted by the neural network by introducing addition, that is, to translate the center coordinate position, as shown in Formula 2. At the same time, multiplication is also introduced to automatically calculate the hyperparameter set of the anchor point, and the feature or prediction box can be scaled, that is, the size of the helmet in the input image is adjusted to display the target box of 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, 1 / 2 to 1 multi-target size of the entire image for scale scaling, as shown in Formula 1. The present invention sets the learning rate value range [0, 0.01], uses the Adam optimizer, and randomly initializes the learning parameters of the model. Then the output predicted label value and the true label value, the output predicted position coordinates and the true position coordinates are input into the cross entropy loss function and GIOU, and the batch gradient descent method is used for back propagation. Finally, the learning parameters of the iterative model are updated 500 times according to the gradient to obtain the optimal detection network model.
[0027]
[0028]
[0029] Where i represents the number of layers of the network model, X i represents the input of the i-th layer, σ represents the activation function SILU, W i represents the weight of the i-th layer, b i is the bias, It is an implicit expression, and the initial value is a random constant tensor. It is used to refine the network model, but since it is a constant tensor, it does not increase the computational complexity. i (X i )+b i represents a convolutional neural network, Indicates the use of implicit expression to scale and resize features or prediction boxes. Indicates that the network model predicts the center coordinate position to be offset, and the weight W i , bias b i With implicit expressions The learning parameters of the iterative model are updated 500 times according to the gradient to obtain the optimal solution, that is, the category prediction loss, confidence prediction loss and bounding box regression loss are minimized.
[0030] For each output Kernel space of the neural network in the training of multiple target sizes where the size of the helmet in the multi-scene input image is 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image, the feature pyramid map fused by the feature map pyramid network is added and multiplied with the implicit representation, as shown in Formula 1 and Formula 2, that is, the Kernel space is translated, rotated, and scaled, and the implicit expression The learning rate is set to a value range of [0, 0.01], and the Adam optimizer is used. The learning parameters of the randomly initialized model are back-propagated using the batch gradient descent method. The learning parameters of the iterative model are updated 500 times according to the gradient to obtain the optimal detection network model.
[0031] When encountering the problem of detecting multiple targets whose sizes in the input image are displayed as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image, the predicted tasks pull each other and cause poor training results. To this end, we scale the multiple targets to the same size by adding implicit expressions in each prediction branch, that is, adjusting the target box whose sizes in the input image are displayed as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image, as shown in Formula 1.
[0032] Module 4: Output Module
[0033] The output module is composed of a convolutional neural network, which is used to integrate the local information with category distinction in the entire network layer. In order to improve the performance of the network, multi-scale feature information is input into the output module for detection, and the output generates a final output vector with category probability, confidence score and bounding box, and uses weighted non-maximum suppression to suppress redundant target boxes, because there is no real value in the actual test prediction. At the same time, only the target box with the highest confidence is retained, and the target box is applied to the final output feature map.
[0034] The training network model specifically includes the following steps:
[0035] Step 1: Input the preprocessed training data set into the neural network model of the target box category classification and predicted box position regression task combining implicit and explicit information for back-propagation training. The learning parameters of the model include weights and bias terms, and the network model is trained from scratch with randomly initialized parameters.
[0036] Step 2: The category prediction loss and confidence prediction loss use the cross entropy loss function, and the bounding box regression loss uses GIoU Loss. GIoU Loss is used to calculate the loss of the prediction box. The formula is as follows:
[0037]
[0038] The predicted box and the true box represent that boxes A, B, and C are the smallest closed shapes, respectively, so that C can contain A and B. Then the ratio of the area in C that does not cover A and B to the total area of C is calculated. After calculating the IoU, the ratio of the area that does not belong to the two boxes to the total area is subtracted.
[0039] Introducing the cross entropy loss function L y′ :
[0040]
[0041]
[0042] Where: y ′ i refers to the i-th value in the actual label, y i Refers to the i-th value in the predicted label. Backpropagation is performed using batch gradient descent according to the loss function to update the learning parameters of the model, including weights and bias terms.
[0043] Step 3: Repeat steps 1 and 2 to continuously iterate and train the network model parameters to obtain the optimal network model.
[0044] The trained model is used to test the helmet wearing detection method.
[0045] Beneficial effects:
[0046] The present invention provides a helmet wearing detection method based on implicit expression, designs a convolution downsampling module, a local feature extraction, a context feature aggregation module, an implicit information guidance module and an output layer network module, and can realize the end-to-end detection task of whether the helmet is worn. The helmet wearing detection network of the present invention not only includes local features but also integrates context feature information in the feature extraction process, thereby improving the detection accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1is a flow chart of a helmet wearing detection method based on implicit expression provided by the present invention;
[0048] Figure 2 is a network structure diagram of a helmet wearing detection method based on implicit expression provided by the present invention; DETAILED DESCRIPTION
[0049] The purpose of the present invention is to provide a helmet wearing detection method based on implicit expression, which can complete the training of the network end-to-end without any post-processing process. The helmet wearing detection method based on implicit expression of the present invention is suitable for the ability to capture helmet targets in all occasions, especially when the size of the helmet in the input image is displayed as 0 to 1 / 8 of the entire image, it can also accurately detect whether the worker is wearing a helmet. It not only includes local features but also integrates context feature information, thereby improving the detection accuracy of the network.
[0050] The present invention will be described in detail below in conjunction with the accompanying drawings. It should be noted that the described embodiments are only intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.
[0051] Figure 1 It is a network flow chart of the helmet wearing detection method based on implicit expression provided by the present invention. Figure 2 : is a network structure diagram of the helmet wearing detection method based on implicit expression provided by the present invention. The helmet wearing detection method based on implicit expression provided by the present invention specifically includes:
[0052] Step 1: Data Preprocessing
[0053] The data preprocessing refers to adaptive image scaling for each input image of a worker wearing a hard hat or not wearing a hard hat, and uniformly scaling the original image to a standard size, i.e., 640×640, and then sending it to the hard hat detection network. Then, Mosaic data enhancement is performed on the data set, i.e., 4 images are used and spliced in a random scaling, random cropping and random arrangement manner. This enhancement method can combine several input images into one.
[0054] Step 2: Build a helmet wearing detection model network
[0055] After preprocessing the input image, a local feature extraction module is used to extract features, especially for marginalized features, to generate feature maps, and then upsampling is performed on the feature map layer to adjust it to the same size, and then the channel dimensions are superimposed two by two, which increases the depth of the network while improving the robustness of the network. The present invention uses an implicit information guidance module to align the features of the feature map pyramid network, and adds and multiplies the feature pyramid map after the fusion of the feature map pyramid network with the implicit representation, that is, the kernel space is translated, rotated and scaled. After that, the context feature aggregation module can be used to aggregate the features in the feature map pyramid network to enhance the global feature information, and then it is passed into the path aggregation network structure in the helmet wearing detection system network, and two feature maps of the same size are spliced in the channel dimension. After two path aggregation network structures, the feature map is re-sampled to the original size. We use the implicit information guidance module to apply to the output module, which is mainly used to predict refinement and learning of different helmet sizes. At the same time, the error term is also modeled with the combined information to guide the network training process.
[0056] Step 3: Training the network model: First, input the preprocessed training data set into the model of the helmet wearing detection network for forward propagation training. The learning rate range is [0, 0.01]. The Adam optimizer is used to randomly initialize the learning parameters of the model. Then, the output predicted label value and the true label value, the output predicted position coordinates and the true position coordinates are input into the cross entropy loss function and GIOU, and the batch gradient descent method is used for back propagation. Finally, the learning parameters of the iterative model are updated 500 times according to the gradient to obtain the optimal detection network model.
[0057] Finally, the trained network model is used to test the target box category classification and prediction box position regression tasks.
[0058] The method of the invention proposes a safety helmet based on implicit expression, which refines the features after context aggregation, highlights the marginalized local feature information, and can still detect in real time while ensuring high accuracy. The introduction of implicit expression enables the algorithm to achieve better results in safety helmet detection, which is beneficial to the ability to capture targets in all occasions and different task situations, especially for workers at a long distance when the helmet appears smaller in the image, and can also exert powerful detection capabilities. Experimental data shows that given 10,000 experimental data of pictures of workers wearing various safety helmets in different scenes, the application of this method can correctly find 9596, with a detection accuracy of 95.9% and a detection speed of 29f / s. Therefore, it also has a high detection accuracy in the context of large-scale complex construction applications.
[0059] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technology can understand and think of changes and substitutions within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A helmet wearing detection method based on implicit expression, Features The following steps are involved: Step 1: Construct a dataset and perform preprocessing. The details are as follows: Use an RGB camera to capture photos of workers wearing or not wearing helmets at the construction site and use annotation tool software to annotate them to create a dataset or use existing public datasets to construct a training dataset. Use Mosaic data enhancement and adaptive image scaling to achieve data amplification and enhancement; Step 2: constructing a neural network for target frame category classification and prediction frame position regression tasks combining implicit and explicit information, wherein the neural network for constructing target frame category classification and prediction frame position regression tasks combining implicit and explicit information includes a local feature extraction module, a context feature aggregation module, an implicit information guidance module, and an output module, wherein the explicit information is feature information directly extracted by the convolutional neural network, namely, the explicit information; The local feature extraction module extracts local features across stages on the channel, and integrates the gradient changes from beginning to end into the feature map. The operation is to split the feature map into two parts, and then the two parts are spliced in the channel dimension through the feature information after the convolution operation; The context feature aggregation module means that three pooling structures with filter sizes of 2×2, 4×4, and 6×6 and a step size of 1 are used for the input image, and then different transmission paths are used to transmit the gradient information, thereby optimizing the repeated gradient information and finally outputting the feature aggregation map; The implicit information guidance module indicates that a set of constant tensors are initialized before the training or prediction phase begins, and the network is refined through implicit information, and explicit knowledge and implicit knowledge are combined and integrated into the training process of the network structure. The implicit expression is applied to the feature map pyramid network feature alignment, prediction refinement and multi-objective learning of the helmet size in the image as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, 1 / 2 to the entire image. In the training of multiple scenes and multiple helmet sizes, in order to align each output Kernel space of the neural network, the feature pyramid map fused by the feature map pyramid network is added and multiplied with the implicit expression, which means that Kernel The implicit information of space being translated, rotated and scaled will be added to the output module for refinement. The specific operations are: by introducing addition, the neural network predicts the offset of the center coordinates; introducing multiplication to automatically calculate the hyperparameter set of the anchor point. When encountering the problem of multiple helmet detection, the size of the helmet is displayed as 0 to 1 / 8 of the entire image and 1 / 2 to the entire size. The output detection used for prediction is also different. Therefore, during the training process, there is a pull between the tasks of predicting different helmet sizes. For this reason, the training is guided by adding implicit expressions in each prediction branch; The output module consists of a convolutional neural network, which is used to integrate the local information with category discrimination in the entire network layer, input the multi-scale feature information into the output module for detection, and generate the final output vector with category probability, confidence score and bounding box. Weighted non-maximum suppression is used to suppress redundant target boxes, and the target box is applied to the output feature map; Step 3: Model training: First, input the preprocessed training data set into the model of the helmet wearing detection network for forward propagation training; then, input the output prediction results and the true label value into the loss function, and use the batch gradient descent method for back propagation; then input the output prediction label value and the true label value, the output prediction position coordinates and the true position coordinates into the cross entropy loss function and GIOU, and use the batch gradient descent method for back propagation. Finally, update the learning parameters of the iterative model 500 times according to the gradient to obtain the optimal detection network model; Step 4: Use the trained classification network model to perform helmet wearing detection method.
2. The helmet wearing detection method according to claim 1, It is characterized in that The preprocessing described in step 1 refers to: the preprocessing process implements data amplification and enhancement by randomly splicing and randomly scaling the input data through Mosaic data enhancement, thereby improving the training speed of the model and the accuracy of the network, and at the same time using the adaptive image scaling method to scale the input image to the input size of the network for normalization.
3. The helmet wearing detection method according to claim 1, It is characterized in that The specific calculation steps of the local feature extraction are as follows: The local feature extraction module extracts local features across stages on the channel by integrating the gradient changes into the feature map from beginning to end. The specific operation is to split the feature map into two parts. One part uses the convolution operation with a convolution kernel of 1×1 and a step size of 1, and the other part inputs the feature map into two convolution operations with a convolution kernel of 1×1 and a step size of 1, a batch norm operation and a SILU activation function, and then uses the residual convolution structure to extract local features across stages, and then uses the convolution operation with a convolution kernel of 1×1 and a step size of 1 to adjust the number of channels, and then this part is spliced with the output result of the previous part in the channel dimension.
4. The helmet wearing detection method according to claim 1, It is characterized in that The context feature aggregation module described in step 1 refers to the use of three filter sizes of 2×2, 4×4, and 6×6 for the input image, and the pooling structure with a step size of 1. The gradient information is transmitted using a parallel transmission path, thereby optimizing the repeated gradient information. Finally, it is combined with the input feature map through a convolution operation with a convolution kernel of 1×1 and a step size of 1, and a feature aggregation map is output, thereby enhancing the feature representation capability. The calculation formula of the multi-scale context aggregation module is as follows: Output=conv(concat(φ max (conv(f));φ max (conv(f));φ max (conv(f));conv(f))+conv(f)) Where f represents the input feature map, conv(f) represents the convolution operation of the input image f with a convolution kernel size of 1*1 and a step size of 1, and concat(x 1 ;x 2 ;…;x n ) represents the input feature map {x 1 ,x 2 ,…,x n }Splicing along the channel direction, φ max The calculation is shown in the formula, C represents the total number of channels of the feature map, c represents the number of channels of the feature map, φ c represents the cth channel of the feature map, (h,w) represents the position in space, φ c (h,w) represents the feature information of the cth channel located at the (h,w) coordinate in the feature map.
5. The helmet wearing detection method according to claim 1, It is characterized in that The implicit information guided module guided training refers to adding a set of constant tensors initialized before the training or prediction phase to the feature map pyramid network kernel space alignment, and applying it to the output module. The implicit information will be added to the output module for refinement. The specific operation is implicit expression. By introducing addition, the offset of the center coordinates of the neural network prediction is made, that is, the center coordinate position is translated, and multiplication is also introduced to automatically calculate the hyperparameter set of the anchor point, and the feature or prediction box is scaled, that is, the size of the helmet in the input image is adjusted to display 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, 1 / 2 to 1 of the entire image. The target box of multiple target sizes is scaled, and the learning rate value range is set to [0, 0.01]. The Adam optimizer is used to randomly initialize the learning parameters of the model, and then the output predicted label value and the true label value, the output predicted position coordinates and the true position coordinates are input into the cross entropy loss function and GIOU, and the batch gradient descent method is used for back propagation. Finally, the learning parameters of the iterative model are updated more than 500 times according to the gradient to obtain the optimal detection network model; Where i represents the number of layers of the network model, X i represents the input of the i-th layer, σ represents the activation function SILU, W i represents the weight of the i-th layer, b i is the bias, It is an implicit expression, and the initial value is a random constant tensor. It is used to refine the network model, but since it is a constant tensor, it does not increase the computational complexity W. i (X i )+b i represents a convolutional neural network, Indicates the use of implicit expression to scale and resize features or prediction boxes. Indicates that the network model predicts the center coordinate position to be offset, and the weight W i , bias b i With implicit expressions Update the learning parameters of the iterative model 500 times according to the gradient to obtain the optimal solution, that is, to minimize the category prediction loss, confidence prediction loss and bounding box regression loss; For each output Kernel space of the neural network in the training of multiple target sizes where the size of the helmet in the multi-scene input image is 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image, the feature pyramid map fused by the feature map pyramid network is added and multiplied with the implicit expression, that is, the Kernel space is translated, rotated and scaled, and the implicit expression The learning rate is set to the range of [0, 0.01], and the Adam optimizer is used. The learning parameters of the randomly initialized model are back-propagated using the batch gradient descent method. The learning parameters of the iterative model are updated 500 times according to the gradient to obtain the optimal detection network model. When encountering the problem of detecting multiple targets whose sizes are displayed as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image in the input image, the predicted tasks pull each other, resulting in poor training results. To this end, an implicit expression is added to each prediction branch to scale the multi-target sizes to the same size, that is, the target boxes whose sizes are displayed as 0 to 1 / 8, 1 / 8 to 1 / 4, 1 / 4 to 1 / 2, and 1 / 2 to 1 of the entire image are scaled.
