Crane recognition device, method and use based on a multi-scale residual detection network

Abstract

The application discloses a crane recognition device and method based on a multi-scale residual detection network and purposes, and relates to the technical field of crane visual detection. The device comprises a detection module, which is used for obtaining a trained multi-scale residual detection network, obtaining a crane image, inputting the crane image into the multi-scale residual detection network, obtaining a crane prediction frame and confidence, and comprises a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network and a prediction end. The multi-scale residual feature extraction network, the receptive field enhancement network, the multi-scale feature fusion network and the prediction end are sequentially connected, and the multi-scale residual feature extraction network is connected with the multi-scale feature fusion network. The method comprises a detection step, obtains a crane image, inputs the crane image into the multi-scale residual detection network, obtains a crane prediction frame and confidence, and further realizes crane detection. The detection module and the like are used to realize good crane detection effect.

Description

Technical Field

[0001] This invention relates to the field of crane visual inspection technology, and in particular to a crane identification device, method and application based on a multi-scale residual detection network. Background Technology

[0002] Currently, municipal and road / bridge construction projects pose numerous hidden dangers to the safe operation of power transmission lines. In particular, the operation of large machinery such as cranes under power lines can easily lead to line contact accidents. Therefore, implementing vision-based crane recognition technology in critical areas of power transmission lines is crucial.

[0003] Machine vision-based target recognition algorithms have been widely used in the field of crane identification. Yan Chunjiang (Yan Chunjiang, Wang Chuang, Fang Hualin, et al. Deep learning-based intrusion detection of power transmission line engineering vehicles [J]. Information Technology, 2018, 42(07):28-33+38.) used PBAS as a moving target detection algorithm to extract suspicious regions between different frames of images and used the VGG Net model for image classification to realize intrusion detection of power transmission channels.

[0004] Liu J (LIU J,HUANG H,ZHANG Y, et al.Deep Learning Based External-force-damage Detection for Power Transmission Line[J].JPh CS,2019,1169(1):012032.) combined Faster R-CNN with the K-means algorithm, using the K-means algorithm to generate anchor boxes, and then using Faster R-CNN to achieve power transmission line intrusion detection.

[0005] Based on the above two documents and existing technical solutions, the inventors have analyzed and found the following technical problems in the existing technical solutions.

[0006] Because crane targets exhibit multi-scale characteristics, and target information is easily lost during image downsampling, the detection effectiveness of the two methods mentioned above is significantly weakened. Under these circumstances, developing a crane detection method with better detection performance to improve the safety guarantee coefficient of power transmission lines is of great practical significance to socio-economic development.

[0007] Existing technical issues and considerations:

[0008] How to solve the technical problem of poor crane inspection results. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a crane identification device, method and application based on a multi-scale residual detection network, so as to solve the technical problem of poor crane detection effect.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A crane recognition device based on a multi-scale residual detection network includes a detection module, which is a program module used to obtain a trained multi-scale residual detection network, obtain a crane image, input the crane image into the multi-scale residual detection network, and obtain the crane prediction box and confidence score. The multi-scale residual detection network includes a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network, and a prediction end. The multi-scale residual feature extraction network, the receptive field enhancement network, the multi-scale feature fusion network, and the prediction end are connected in sequence, and the multi-scale residual feature extraction network and the multi-scale feature fusion network are connected.

[0011] A further technical solution includes a camera, a communication device, and a processor for acquiring crane images, wherein the camera is connected to and communicates with the communication device, and the communication device is connected to and communicates with the processor.

[0012] A further technical solution is that the detection module is also used to make the crane image a preprocessed crane image.

[0013] A further technical solution involves a detection module that also sends the crane prediction frame and confidence level to the outside world.

[0014] A crane recognition method based on a multi-scale residual detection network includes the following steps: obtaining a crane image, inputting the crane image into a multi-scale residual detection network, obtaining the crane prediction box and confidence score, and thus realizing crane detection.

[0015] A further technical solution is to preprocess the crane image and obtain the preprocessed crane image in the detection step, and then input the preprocessed crane image into a multi-scale residual detection network.

[0016] A further technical solution involves sending the crane prediction frame and confidence level to external parties and promptly notifying management personnel during the detection process.

[0017] A crane identification device based on a multi-scale residual detection network includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the corresponding steps in the above method.

[0018] A crane identification device based on a multi-scale residual detection network includes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the corresponding steps in the above method.

[0019] Application of a multi-scale residual detection network: using a multi-scale residual detection network for crane inspection.

[0020] The beneficial effects of adopting the above technical solution are as follows:

[0021] A crane recognition device based on a multi-scale residual detection network includes a detection module, which is a program module used to obtain a trained multi-scale residual detection network, acquire a crane image, input the crane image into the multi-scale residual detection network, and obtain the crane prediction bounding box and confidence score. The multi-scale residual detection network includes a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network, and a prediction terminal. These components are sequentially connected, with the multi-scale residual feature extraction network and the multi-scale feature fusion network connected together. This technical solution achieves good crane detection results through the detection module and other features.

[0022] A crane recognition method based on a multi-scale residual detection network includes detection steps: obtaining a crane image; inputting the crane image into the multi-scale residual detection network to obtain the crane prediction box and confidence score, thereby achieving crane detection. This technical solution achieves good crane detection results through its detection steps.

[0023] A crane identification device based on a multi-scale residual detection network includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the corresponding steps in the above method, achieving good crane detection results.

[0024] A crane identification device based on a multi-scale residual detection network includes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the corresponding steps in the above method, achieving good crane detection results.

[0025] An application of a multi-scale residual detection network is described, which is used for crane inspection to achieve good crane inspection results.

[0026] See the detailed implementation section for further description. Attached Figure Description

[0027] Figure 1 This is a principle block diagram of Embodiment 1 of the present invention;

[0028] Figure 2 yes Figure 1 Block diagram of the Stage1 submodule;

[0029] Figure 3 yes Figure 1Block diagram of the Stage2 submodule;

[0030] Figure 4 yes Figure 1 Block diagram of the receptive field enhancement network;

[0031] Figure 5 yes Figure 1 A schematic diagram of the principle of a multi-scale feature fusion network;

[0032] Figure 6 This is a flowchart illustrating the implementation of the present invention;

[0033] Figure 7 This is a diagram showing the detection results of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this application or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0035] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0036] Example 1:

[0037] like Figures 1-5 As shown, the present invention discloses a crane identification device based on a multi-scale residual detection network, including a camera for obtaining crane images, a communication device and a processor, and a detection module. The camera is connected to and communicates with the communication device, and the communication device is connected to and communicates with the processor.

[0038] The detection module is a program module used by the processor to obtain the trained multi-scale residual detection network, the processor to obtain the crane image sent by the camera, the processor to preprocess the crane image to obtain the preprocessed crane image, the processor to input the preprocessed crane image into the multi-scale residual detection network to obtain the crane prediction box and confidence score, and the processor to send the crane prediction box and confidence score to the outside world through the communication device.

[0039] like Figure 1As shown, the multi-scale residual detection network includes a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network, and a prediction end. The multi-scale residual feature extraction network, the receptive field enhancement network, the multi-scale feature fusion network, and the prediction end are connected in sequence, and the multi-scale residual feature extraction network is connected to the multi-scale feature fusion network.

[0040] like Figure 2 As shown in the figure, the structure of the Stage1 submodule is as follows.

[0041] like Figure 3 As shown in the figure, the structure of the Stage2 submodule is as follows.

[0042] like Figure 4 As shown in the figure, the structure of the receptive field enhancement network is as follows.

[0043] like Figure 5 As shown in the figure, the structure of the multi-scale feature fusion network is as follows.

[0044] Among them, the multi-scale residual feature extraction network, receptive field enhancement network, multi-scale feature fusion network and prediction end are existing technologies. The invention lies in the connection relationship between the multi-scale residual feature extraction network, receptive field enhancement network, multi-scale feature fusion network and prediction end.

[0045] Example 2:

[0046] This invention discloses a crane identification method based on a multi-scale residual detection network, including the following detection steps:

[0047] The system acquires crane images, preprocesses the crane images, and obtains preprocessed crane images. The preprocessed crane images are then input into a multi-scale residual detection network to obtain crane prediction boxes and confidence scores, thereby achieving crane detection. The crane prediction boxes and confidence scores are then sent externally and promptly notified to management personnel.

[0048] Example 3:

[0049] The present invention discloses a crane identification device based on a multi-scale residual detection network, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the steps of Embodiment 2.

[0050] Example 4:

[0051] The present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in Embodiment 2.

[0052] Example 5:

[0053] This invention discloses the application of a multi-scale residual detection network for crane detection.

[0054] Technical contributions of this application:

[0055] This application provides a crane recognition method based on a multi-scale residual network, comprising three steps: the first step is to acquire and preprocess crane images; the second step is to build a multi-scale residual detection network model; and the third step is to train the model and perform crane target detection. The multi-scale residual detection network model includes a multi-scale residual feature extraction network module, a receptive field enhancement network module, a multi-scale feature fusion network module, and a prediction module connected in sequence. The multi-scale residual feature extraction network module is also connected to the multi-scale feature fusion network module.

[0056] Technical solution description:

[0057] like Figure 6 As shown, this application provides a crane recognition method based on a multi-scale residual network, mainly applied to the detection of crane targets. The method includes the following steps:

[0058] The first step is to acquire and preprocess the crane image:

[0059] 1-1 Image Acquisition: 1895 crane images were collected using drone inspections as the original images for crane inspection, and the size of each image was uniformly adjusted to 2592X1944; the color, visibility, and other aspects of the crane varied in each image.

[0060] 1-2 Image Annotation: Based on step 1-1, use annotation software to manually select and annotate the images from step 1-1, and add Chinese crane labels; modify all existing crane images with annotation information, including adjusting the size of the annotation boxes, changing Chinese labels to English labels, and deleting abnormally captured images, thereby obtaining an XML file (automatically generated by the annotation software).

[0061] 1-3 Dataset Creation: Based on steps 1-2, the obtained XML annotation files are converted into TXT and JSON files to obtain the dataset.

[0062] 1-4 Test Sample Set Creation: Based on steps 1-3, randomly select 20% of the images from the obtained dataset (randomization method is not limited) as the test sample set.

[0063] 1-5 Creating Training and Validation Sample Sets: Based on steps 1-4, extract the remaining dataset as the training and validation sample sets.

[0064] The second step is to build a multi-scale residual detection network model:

[0065] like Figure 1 The diagram shown is a structural diagram of a multi-scale residual detection network model. The multi-scale residual detection network model includes a multi-scale residual feature extraction network module, a receptive field enhancement network module, a multi-scale feature fusion network module, and a prediction module connected in sequence. The multi-scale residual feature extraction network module is also connected to the multi-scale feature fusion network module.

[0066] 2-1 Multi-scale Residual Feature Extraction Network Module: This module consists of an initial CBM convolutional structure, Stage 1 sub-module, Stage 2 sub-module, Stage 3 sub-module, Stage 4 sub-module, and Stage 5 sub-module, connected sequentially. The CBM convolutional structure is composed of a convolutional layer (Conv), a batch normalization operation (BN), and an activation function (Mish), connected sequentially. The kernel size of the convolutional layer in the initial CBM convolutional structure is 3x3. After the crane image is resized to 608x608, it is fed into the network. After passing through the initial CBM convolutional structure, the resulting feature map size becomes 608x608x32 and is input into the Stage 1 sub-module, and then sequentially passes through the Stage 2, Stage 3, Stage 4, and Stage 5 sub-modules. The network first processes a block of data and sequentially obtains feature maps S1, S2, S3, S4, and S5, with corresponding dimensions of 304x304x64, 152x152x128, 76x76x256, 38x38x512, and 19x19x1024, respectively. Feature map S5 is then input into the receptive field enhancement network module to obtain feature map S6. Feature maps S3, S4, and S6 are then input into the multi-scale feature fusion network module. After S5 enters the receptive field enhancement network, feature map S6 is output and input into the multi-scale feature fusion network. Finally, the multi-scale feature fusion network module outputs three feature maps F3, F4, and F5. Feature maps F3, F4, and F5 are each processed through a convolution operation and then fed into the prediction end to output result maps Y1, Y2, and Y3, respectively.

[0067] like Figure 2 As shown, the Stage1 submodule consists of a CBM convolutional structure with a kernel size of 3x3, a unit sub-block, and a Residual structure connected in sequence.

[0068] The structure of the unit sub-block includes a CBM convolutional structure with a kernel size of 1x1, an MSCBM convolutional structure, and a residual structure connected in sequence.

[0069] The MSCBM convolutional structure is a hierarchical residual structure. Its structural principle is as follows: the original single input channel is divided into four sub-channels: N1, N2, N3, and N4. Each of the N2, N3, and N4 sub-channels contains a convolutional layer with a kernel size of 3x3. The channel before the convolutional layer in N3 is connected to the output channel of the N2 sub-channel, and the channel before the convolutional layer in N4 is connected to the output channel of N3. The four sub-channels N1, N2, N3, and N4 are finally concatenated into a single output channel through a concat operation and then connected to a CBM convolutional structure with a kernel size of 1x1. Finally, the input to the unit sub-block and the output of the MSCBM convolutional structure are concatenated once.

[0070] The Residual structure consists of two CBM convolutional structures with a kernel size of 3x3 connected in sequence. The output after passing through the two CBM convolutional structures is added to the original input entering the Residual structure by one pixel.

[0071] like Figure 3 As shown, the structure of the Stage2 submodule includes a CBM convolutional structure with a kernel size of 3x3 connected in sequence, two repeating units, and a Residual structure; the structure of the repeating units is the same as the structure of the unit sub-blocks in the Stage2 module; the Residual structures in both the Stage2 and Stage1 submodules are the same.

[0072] The Stage3 submodule is formed by adding 6 repeating units after the repeating units in the Stage2 submodule.

[0073] The structure of the Stage4 submodule is the same as that of the Stage3 submodule.

[0074] The Stage5 submodule is formed by adding two more repeating units after the repeating units in the Stage2 submodule.

[0075] 2-2 Receptive Field Enhancement Network Module:

[0076] like Figure 4 As shown, the receptive field enhancement network structure includes an input terminal, three dilated convolutional layers of size 3x3 with dilation rates of 1, 3, and 5, and an SE block module. The output channel of the input terminal is divided into three sub-channels N5, N6, and N7, which are connected to the three dilated convolutional layers respectively. The output channels of the three dilated convolutional layers are restored to a single output channel through a Concat concatenation operation and then connected to the SE block module after a 1x1 convolution operation.

[0077] The SE block module consists of a residual structure. The input to the SE block module passes through the residual structure, global pooling, a fully connected layer (FC), a ReLU activation function, another fully connected layer (FC), and a Sigmoid activation function in sequence. The output obtained is then combined with the output of the residual structure and subjected to a scaling operation. Finally, the output after the scaling operation is concatted with the input to the SE block module. The residual structures in the SE block and Stage 1 modules are the same.

[0078] 2-3 Multi-scale Feature Fusion Network Modules:

[0079] like Figure 5 As shown, the structure of the multi-scale feature fusion network module includes four concat operations; feature maps S6 and S4 are concatted together to obtain feature map P4, feature maps P4 and S3 are concatted together to obtain feature map F3, feature maps P4 and F3 are concatted together to obtain feature map F4, and feature maps S6 and F4 are concatted together to obtain feature map F5.

[0080] 2-4 Prediction Module:

[0081] First, the K-Means algorithm is used to perform cluster analysis on the labeled training data, and three sets of prior boxes of different sizes are obtained for the three feature maps with sizes of 19x19, 38x38 and 76x76 respectively.

[0082] As shown in Table 1, the specific prior box details are as follows.

[0083] Table 1: Details of the prior bounding box

[0084]

[0085] Then, the feature maps F3, F4 and F5 are divided into cells of size 19x19, 38x38 and 76x76 respectively, and a set of prior boxes adapted to the size of each cell is set to predict the bounding box of the target.

[0086] The bounding box prediction process is as follows: First, a prior box with a width of p is pre-set. w , height is p h Its center point coordinates are (c x c y ); by predicting the offset t of the crane target position coordinates. x , t y , t w, t h This yields the actual position of the crane target; then, the predicted target's bounding box is calculated using the predicted offset, and its center point coordinates (b x b y ), with a width of b w The height is b h The calculation formula is:

[0087] b x =δ(t) x )+c x (1)

[0088] b y =δ(t) y )+c y (2)

[0089]

[0090]

[0091] Where δ represents the Sigmoid function.

[0092] The third step is to train the model and perform crane target detection:

[0093] 3-1 Parameter Initialization: Load the pre-trained weight file and initialize momentum, batch size, weight values, and iteration period.

[0094] 3-2 Start training: Input the training sample set into the network model after initializing the parameters in step 3-1.

[0095] 3-3 Update parameters: Compare the parameters of the predicted target box and the real target box output in step 3-2, return the loss function loss to the network model trained in step 3-2, and use the gradient descent algorithm to pass the gradient loss to the parameters of each convolution kernel to update the trainable parameters.

[0096] 3-4 Stop training: Repeat training step 3-3. When the number of training cycles is greater than 1000, stop training to obtain the final multi-scale residual detection network model.

[0097] 3-5 Testing the crane detection effect: Input the test sample set data into the final multi-scale residual detection network model to obtain the category label, the predicted crane bounding box in the image, and the confidence score.

[0098] 3-6 Complete the inspection of the crane target.

[0099] The technical solution of this application is implemented based on the PyTorch (1.5.1) deep learning framework, and the computer performance used is: CPU is a Core i7 series processor, 64GB memory, and graphics card is NVIDIA TITAN XP.

[0100] The technical solution of this application uses mean accuracy (MAP) and detection speed (fps) to measure the performance of the network model.

[0101] Table 2 shows the comparison data for different models:

[0102] As shown in Table 2, on the one hand, the YOLOv5x model has the highest MAP value among the other models, reaching 69.9%, but it is still 1.9 percentage points lower than the model of this application. Moreover, the detection speed of the model of this application is 74 fps higher than that of the YOLOv5x model. On the other hand, the YOLOv5s model has the highest fps value among the other models, with a value of 112 fps, which is 9 fps lower than that of the model of this application.

[0103] Table 2: MAP and FPS values ​​for different models

[0104]

[0105] like Figure 7 The image shown is a diagram illustrating the testing results of the technical solution presented in this application.

[0106] After this application was kept confidential for a period of time, the beneficial aspects reported by the on-site technical personnel were:

[0107] The technical solution of this application, while ensuring that the MAP index reaches more than 70%, has a high detection speed of 121fps, which can simultaneously meet the requirements of accuracy and real-time speed, effectively improving the detection effect of crane inspection, and can be widely used in the inspection of cranes near high-voltage power lines.

[0108] A crane recognition device based on a multi-scale residual detection network includes a detection module, which is a program module used to obtain a trained multi-scale residual detection network, acquire a crane image, input the crane image into the multi-scale residual detection network, and obtain a crane prediction box and confidence score. The multi-scale residual detection network includes a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network, and a prediction terminal. The multi-scale residual feature extraction network, the receptive field enhancement network, the multi-scale feature fusion network, and the prediction terminal are connected in sequence. The multi-scale residual feature extraction network and the multi-scale feature fusion network are connected. Through the detection module, etc., the device achieves good crane detection results.

[0109] A crane recognition method based on a multi-scale residual detection network includes detection steps: obtaining a crane image, inputting the crane image into a multi-scale residual detection network to obtain a crane prediction box and confidence score, thereby realizing crane detection. Through these detection steps, the method achieves good crane detection results.

[0110] A crane identification device based on a multi-scale residual detection network includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the corresponding steps in the above method, achieving good crane detection results.

[0111] A crane identification device based on a multi-scale residual detection network includes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the corresponding steps in the above method, achieving good crane detection results.

[0112] An application of a multi-scale residual detection network is described, which is used for crane inspection to achieve good crane inspection results.

[0113] Currently, the technical solution of this invention has undergone pilot testing, which is a small-scale trial of the product before large-scale mass production. After the pilot testing was completed, a user survey was conducted on a small scale, and the survey results showed that user satisfaction was high. Now, preparations have begun for the formal production and industrialization of the product (including intellectual property risk warning surveys).

Claims

1. A crane identification device based on a multi-scale residual detection network, characterized in that: The system includes a detection module, which is a program module used to obtain a trained multi-scale residual detection network, obtain a crane image, input the crane image into the multi-scale residual detection network, and obtain the crane prediction box and confidence score. The multi-scale residual detection network includes a multi-scale residual feature extraction network, a receptive field enhancement network, a multi-scale feature fusion network, and a prediction terminal. The multi-scale residual feature extraction network, the receptive field enhancement network, the multi-scale feature fusion network, and the prediction terminal are connected in sequence, and the multi-scale residual feature extraction network and the multi-scale feature fusion network are connected. The multi-scale residual feature extraction network includes an initial CBM convolutional structure, Stage1 submodule, Stage2 submodule, Stage3 submodule, Stage4 submodule, and Stage5 submodule connected in sequence. Stage3 submodule is connected to the multi-scale feature fusion network, and Stage4 submodule is connected to the multi-scale feature fusion network. The CBM convolutional structure consists of a convolutional layer Conv, a batch normalization operation BN, and an activation function Mish, which are connected sequentially. The kernel size in the initial CBM convolutional structure is 3x3. The Stage1 submodule includes a CBM convolutional structure with a kernel size of 3x3, a unit sub-block, and a Residual structure connected in sequence. The unit sub-block includes a CBM convolutional structure with a kernel size of 1x1, an MSCBM convolutional structure, and a Residual structure connected in sequence. The MSCBM convolutional structure is a hierarchical residual structure. The Residual structure includes two CBM convolutional structures with a kernel size of 3x3 connected in sequence. The Stage2 submodule includes a CBM convolutional structure with a kernel size of 3x3, a first repeating unit, a second repeating unit, and a residual structure connected in sequence; the structure of the repeating unit is the same as the structure of the unit sub-block; The structure of the Stage3 submodule is formed by adding six repeating units connected in sequence after the second repeating unit in the structure of the Stage2 submodule. The structure of the Stage4 submodule is the same as that of the Stage3 submodule; The structure of the Stage5 submodule is formed by adding two repeating units connected in sequence after the second repeating unit in the structure of the Stage2 submodule. The receptive field enhancement network includes an input terminal, three dilated convolutional layers of size 3x3 with dilation rates of 1, 3, and 5, and an SE block module. The input terminal is connected to each dilated convolutional layer. The output channels of all dilated convolutional layers are concatenated into a single output channel by a Concat operation and then connected to the SE block module after a 1x1 convolution operation. The K-Means algorithm is used for cluster analysis to obtain prior boxes in the prediction process.

2. The crane identification device based on a multi-scale residual detection network according to claim 1, characterized in that: It also includes a camera, a communication device, and a processor for acquiring images of the crane, wherein the camera is connected to and communicates with the communication device, and the communication device is connected to and communicates with the processor.

3. The crane identification device based on a multi-scale residual detection network according to claim 1, characterized in that: The detection module is also used to ensure that the crane image is a pre-processed crane image.

4. The crane identification device based on a multi-scale residual detection network according to claim 2, characterized in that: The detection module is also used to send the crane prediction box and confidence score to external systems.

5. A crane identification method based on a multi-scale residual detection network, characterized in that: The crane recognition device based on the multi-scale residual detection network as described in claim 1 includes a detection step: obtaining a crane image, inputting the crane image into the multi-scale residual detection network, obtaining the crane prediction box and confidence level, and thus realizing crane detection.

6. The crane identification method based on a multi-scale residual detection network according to claim 5, characterized in that: In the detection step, the crane image is preprocessed to obtain the preprocessed crane image, and the preprocessed crane image is input into the multi-scale residual detection network.

7. The crane identification method based on a multi-scale residual detection network according to claim 6, characterized in that: During the detection process, the crane prediction frame and confidence level are sent to external parties and the management personnel are notified in a timely manner.

8. A crane identification device based on a multi-scale residual detection network, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the corresponding steps in the method of any one of claims 5 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the corresponding steps in the method of any one of claims 5 to 7.

10. An application of a multi-scale residual detection network, characterized in that: Based on the crane identification device based on the multi-scale residual detection network as described in claim 1, the multi-scale residual detection network is used for crane detection.

Images