Method for automatically checking container damage based on pattern recognition

Abstract

The invention provides a method for automatically checking container damage based on pattern recognition, which comprises the following steps of: S1, acquiring a container body image of each surface of any container body, preprocessing to obtain a preprocessed container body image, and dividing the preprocessed container body image into a training set and a test set; s2, building a container damage checking network which comprises a backbone module, a feature fusion module and a damage identification module; s3, training the container damage inspection network by using the training set to obtain a container damage identification image; s4, training the container damage inspection network for multiple rounds according to the step S3 to obtain a container damage inspection network model; and S5, testing the test set through the container damage checking network model, and outputting a container damage identification image corresponding to each preprocessed box body image as an automatic container damage checking result. The method has the beneficial effects that the surface condition of the container body can be quickly and accurately checked in an unmanned operation mode to obtain a damage identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of container damage identification. Specifically, it relates to an automatic inspection method for container damage based on pattern recognition. Background Art

[0002] With the development of the transportation and water transportation industries, port operations (such as loading and unloading operations) are becoming increasingly frequent. To ensure the safety of cargo transportation, goods are usually loaded into containers for transportation. Among them, a container is a large loading container with a certain strength, stiffness, and specifications and is specifically for repeated use.

[0003] At current maritime transportation terminals, the problem of container damage inspection relies on manual judgment and visual inspection, resulting in low efficiency; some terminals and research institutes have adopted two-stage object detection technologies represented by Mask-RCNN to detect container damage in order to replace manual detection; however, whether it is the interactive recognition of images or videos, accurate identification of specific damage categories is required in the field of container damage inspection, which requires providing accurate and reliable information.

[0004] Its main purpose is to detect, locate, and classify specific targets from static pictures or videos. In actual situations, the damage classification of container bodies often varies, the amount of data is small, and it also brings higher recognition challenges in different weather and light intensity conditions. There is an urgent need for a new method for inspecting the damage of the container body surface at the container terminal shore, aiming to quickly and accurately check the surface condition of the container body in an unmanned operation mode to obtain the damage recognition result. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to quickly and accurately check the surface condition of the container body in an unmanned operation mode to obtain the damage recognition result. To overcome the defects of the above prior art (or related technologies), the present invention provides an automatic inspection method for container damage based on pattern recognition.

[0006] The present invention provides an automatic inspection method for container damage based on pattern recognition, including: Step S1, for any container body, obtain the container body image of each surface of the container body, preprocess each container body image to obtain the corresponding preprocessed container body image, and divide it into a training set and a test set; Step S2, use a deep learning framework to build a container damage inspection network. The container damage inspection network includes a backbone module, a feature fusion module, and a damage recognition module. The backbone module receives the input image and outputs a feature map 、feature map and feature map , and through the feature fusion module, for the feature map and the feature map and the feature map perform pixel weight adjustment operation, channel splicing operation and convolution operation to obtain a feature map feature map and feature map , and the damaged recognition module processes the feature map , the feature map and the feature map to obtain a container damaged recognition image; Step S3, use the training set to train the container damaged inspection network to obtain the container damaged recognition image corresponding to each preprocessed box image in the training set; Step S4, train the container damaged inspection network multiple rounds according to the step S3 to obtain a container damaged inspection network model; Step S5, test each preprocessed box image in the test set through the container damaged inspection network model, and output the container damaged recognition image corresponding to each preprocessed box image as the automatic inspection result of container damage.

[0007] Compared with the prior art, a method for automatically inspecting container damage based on pattern recognition in this application has the following advantages: In this application, the preprocessing of the box images is performed through step S1, the construction of the container damaged inspection network is performed through step S2, the training of the container damaged inspection network is performed through step S3, the generation of the container damaged inspection network model is performed through step S4, and the testing of the container damaged inspection network model and the generation of the automatic inspection result of container damage are performed through step S5. Compared with manual evaluation, the standard is unified and the efficiency is improved; compared with classical image processing methods and existing inspection technologies, it is more in line with the actual situation, has higher accuracy and lower missed inspection rate, enabling the rapid and accurate inspection of the surface condition of the container body in an unmanned operation mode.

[0008] In a possible implementation manner, in the step S1, the process of preprocessing each box image includes: Perform mirror flipping, rotation, scaling or translation operations on each box image respectively to obtain a plurality of preprocessed box images.

[0009] In a possible implementation manner, the backbone module includes a first CBS layer, a second CBS layer, a first C2f layer, a third CBS layer, a second C2f layer, a fourth CBS layer, a first DC2f layer, a fifth CBS layer and a second DC2f layer connected in sequence. The input image is received by the first CBS layer and a feature map is output Receive the feature map through the second CBS layer and output the feature map Receive the feature map through the first C2f layer and output the feature map Receive the feature map through the third CBS layer and output the feature map Receive the feature map through the second C2f layer and respectively output the feature map and the feature map Receive the feature map through the fourth CBS layer and output the feature map Receive the feature map through the first DC2f layer and respectively output the feature map and the feature map Receive the feature map through the fifth CBS layer and output the feature map Receive the feature map through the second DC2f layer and output the feature map The size of the input image is 640x640x3, and the feature map has a size of 320x320x64, and the feature map has a size of 160x160x128, and the feature map has a size of 160x160x128, and the feature map has a size of 80x80x256, and the feature map has a size of 80x80x256, and the feature map has a size of 80x80x256, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 20x20x512, and the feature map has a size of 20x20x512.

[0010] In a possible implementation manner, both the first DC2f layer and the second DC2f layer include a first DCBS layer, a Split layer, a plurality of Bottleneck layers, a first concat layer, and a second DCBS layer that are connected in sequence, and the Split layer is respectively connected to each of the Bottleneck layers.

[0011] In a possible implementation manner, both the first DCBS layer and the second DCBS layer include a deformable convolutional layer, a BN layer, and a SiLU layer that are connected in sequence.

[0012] In a possible implementation manner, each of the Bottleneck layers includes two CBS layers and an add layer that are connected in sequence, and the add layer is connected to the first CSB layer.

[0013] In a possible implementation manner, the feature fusion module includes an SPPF-SimAM layer, a first CARAFE layer, a second concat layer, a third C2f layer, a second CARAFE layer, a third concat layer, a fourth C2f layer, a sixth CBS layer, a fourth concat layer, a fifth C2f layer, a seventh CBS layer, a fifth concat layer, and a sixth C2f layer that are connected in sequence. The SPPF-SimAM layer is also connected to the fifth concat layer, and the third C2f layer is also connected to the fourth concat layer. The feature map is received through the SPPF-SimAM layer and feature maps are respectively output and feature map , the feature map is received through the first CARAFE layer and feature map is output , the feature map is received through the second concat layer and the feature map and feature map is output , the feature map is received through the third C2f layer and feature maps are respectively output and feature map , the feature map is received through the second CARAFE layer and feature map is output , the feature map is received through the third concat layer and the feature map and feature map is output , the feature map is received through the fourth C2f layer and feature map is output and the feature map , the feature map is received through the sixth CBS layer and feature map is output , the feature map is received through the fourth concat layer and the feature map and feature map is output , the feature map is received through the fifth C2f layer And output the feature map And the said feature map , receive the said feature map through the seventh CBS layer And output the feature map , receive the said feature map through the fifth concat layer And the said feature map And output the said feature map .

[0014] In a possible implementation manner, the size of the said feature map is 20x20x512, the size of the said feature map is 20x20x512, the size of the said feature map is 40x40x512, the size of the said feature map is 40x40x1024, the size of the said feature map is 40x40x512, the size of the said feature map is 40x40x512, the size of the said feature map is 80x80x512, the size of the said feature map is 80x80x768, the size of the said feature map is 80x80x256, the size of the said feature map is 80x80x256, the size of the said feature map is 40x40x256, the size of the said feature map is 40x40x768, the size of the said feature map is 40x40x512, the size of the said feature map is 40x40x512, the size of the said feature map is 20x20x512, the size of the said feature map is 20x20x1024, the size of the said feature map is 20x20x512.

[0015] In a possible implementation manner, the damaged recognition module includes three Detect layers, and respectively receive the said feature map , the said feature map , the said feature map through the three Detect layers to perform defect recognition and marking of the prediction box to obtain the container damaged recognition image.

[0016] In a possible implementation, each of the Detect layers includes two branches, and each branch only includes two CBS layers, a Conv layer, and a Bbox Loss layer connected in sequence. Description of the Drawings

[0017] Figure 1 It is a flowchart of the steps of the present invention; Figure 2 It is a schematic structural diagram of the container damage inspection network of the present invention; Figure 3 It is a schematic structural diagram of the DC2f layer of the present invention; Figure 4 It is a schematic structural diagram of the DCBS layer of the present invention; Figure 5 It is a schematic structural diagram of the SPPF-SimAM layer of the present invention. Detailed Embodiments

[0018] First of all, those skilled in the art should understand that these embodiments are only used to explain the technical principles of the embodiments of this application, and are not intended to limit the protection scope of the embodiments of this application. Those skilled in the art can adjust it according to needs to adapt to specific application scenarios.

[0019] The following further describes this application in detail with reference to the drawings and specific embodiments.

[0020] See Figure 1 , the embodiments of this application disclose an automatic inspection method for container damage based on pattern recognition, including: Step S1, for any container body, obtain the container body images of each surface of the container body, preprocess each container body image to obtain the corresponding preprocessed container body image, and divide it into a training set and a test set; Step S2, use a deep learning framework to build a container damage inspection network. The container damage inspection network includes a backbone module, a feature fusion module, and a damage recognition module. The backbone module receives the input image and outputs feature map , feature map and feature map , through the feature fusion module, perform pixel weight adjustment operation, channel splicing operation, and convolution operation on feature map , feature map and feature map to obtain feature map , feature map and feature map , through the damage recognition module, for feature map , feature map and feature map Process to obtain the container damage recognition image; Step S3: Use the training set to train the container damage inspection network to obtain the container damage recognition image corresponding to each preprocessed container body image in the training set; Step S4: Train the container damage inspection network multiple rounds according to Step S3 to obtain the container damage inspection network model; Step S5: Test each preprocessed container body image in the test set through the container damage inspection network model, and output the container damage recognition image corresponding to each preprocessed container body image as the automatic inspection result of container damage.

[0021] This application explores the research status of target detection algorithms at home and abroad. Currently, there are few relevant studies on container damage detection, and there is no mature research result in relying on vision technology to detect various types of container body damage. At the same time, it is explored that the implementation maturity of leading domestic terminals in container damage inspection is still at a relatively low level, and relevant patents and papers are scarce. The research value and application prospect of this application are forward-looking, novel, and creative in the field of container terminals.

[0022] This application proposes a new inspection method for automatically inspecting the surface damage of container bodies at the quay of container terminals. Compared with manual judgment, it unifies the standards and improves the efficiency; compared with classical image processing methods, the existing inspection technologies and methods are more in line with the actual situation, have higher accuracy, and lower missed inspection rates, providing a solution for the maturity of the container automatic damage inspection module, specifically as follows: 1. This application is improved based on the YOLOv8 network structure. In the improvement of the feature extraction network, the idea of optimizing the original backbone network is adopted, and the lightweight backbone network MobileNetv4 is introduced with the focus on reducing the size of the network model to improve the detection speed of the container damage inspection model; 2. In the neck part of the container damage inspection network, with the focus on reducing model false detection and missed detection situations, a CSPStage structure is designed to replace the original C2 module, and at the same time, the self-attention mechanism is combined to improve the feature extraction efficiency; 3. Aiming at the problems of many damage types and few data in the dataset, data augmentation technologies such as mirror flipping, rotation, scaling, and translation are used to increase the sample data of the training set. At the same time, by inferring to obtain the recognized pictures, semi-supervised annotation is used to increase the sample data volume to meet the dataset quantity requirements for damage learning and inference in different cameras and different pictures.

[0023] This application is based on the YOLOv8l model. First, a dynamic SPPF-SimAM (SPPFS) optimization structure is designed to replace the original SPPF structure, dynamically adjusting the weights of each pixel to enhance the container damage inspection model's ability to capture details of container damage images. At the same time, a DCBS (Deformable convolution BatchNormalization SiLU) structure is introduced to reconstruct the C2f structure into DC2f, optimizing the relatively rigid problem of C2f in the process of container damage detection. By replacing the upsampling operator with a lightweight CARAFE operator, it can be further lightweight and fast, and can adapt to container damage situations of different shapes, sizes, and materials, flexibly adjusting according to the damage type and reducing the limitations on the detection effect. Finally, aiming at the problems of many damage types and few data in the dataset, data augmentation techniques such as mirror flipping, rotation, scaling, and translation are used to increase the sample data of the training set. At the same time, by inferring to obtain the recognized pictures, semi-supervised annotation is used to increase the sample data volume to meet the dataset quantity requirements for damage learning and inference in different cameras and different scenarios.

[0024] See Figure 2 , the backbone module includes a first CBS layer, a second CBS layer, a first C2f layer, a third CBS layer, a second C2f layer, a fourth CBS layer, a first DC2f layer, a fifth CBS layer, and a second DC2f layer connected in sequence. The input image is received by the first CBS layer and a feature map is output , the feature map is received by the second CBS layer and a feature map is output , the feature map is received by the first C2f layer and a feature map is output , the feature map is received by the third CBS layer and a feature map is output , the feature map is received by the second C2f layer and feature maps are respectively output and feature map , the feature map is received by the fourth CBS layer and a feature map is output , the feature map is received by the first DC2f layer and feature maps are respectively output and feature map , the feature map is received by the fifth CBS layer and a feature map is output , the feature map is received by the second DC2f layer and a feature map is output ; The feature fusion module includes an SPPF-SimAM layer, a first CARAFE layer, a second concat layer, a third C2f layer, a second CARAFE layer, a third concat layer, a fourth C2f layer, a sixth CBS layer, a fourth concat layer, a fifth C2f layer, a seventh CBS layer, a fifth concat layer, and a sixth C2f layer connected in sequence. The SPPF-SimAM layer is also connected to the fifth concat layer, and the third C2f layer is also connected to the fourth concat layer. The feature map is received through the SPPF-SimAM layer and respectively outputs a feature map and a feature map , receives the feature map through the first CARAFE layer and outputs a feature map , receives the feature map and the feature map through the second concat layer and outputs a feature map , receives the feature map through the third C2f layer and respectively outputs a feature map and a feature map , receives the feature map through the second CARAFE layer and outputs a feature map , receives the feature map and the feature map through the third concat layer and outputs a feature map , receives the feature map through the fourth C2f layer and outputs a feature map and a feature map , receives the feature map through the sixth CBS layer and outputs a feature map , receives the feature map and the feature map through the fourth concat layer and outputs a feature map , receives the feature map through the fifth C2f layer and outputs a feature map and a feature map , receives the feature map through the seventh CBS layer and outputs a feature map , receives the feature map and the feature map through the fifth concat layer and outputs a feature map ; The damage recognition module includes three Detect layers, and respectively receives the feature maps , the feature map , the feature map Defect recognition of the prediction box is performed and marked to obtain the container damage recognition image.

[0025] Continue to refer to Figure 2 , the size of the input image is 640x640x3, and the feature map has a size of 320x320x64, and the feature map has a size of 160x160x128, and the feature map has a size of 160x160x128, and the feature map has a size of 80x80x256, and the feature map has a size of 80x80x256, and the feature map has a size of 80x80x256, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 20x20x512, and the feature map has a size of 20x20x512; the feature map has a size of 20x20x512, and the feature map has a size of 20x20x512, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x1024, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 80x80x512, and the feature map has a size of 80x80x768, and the feature map has a size of 80x80x256, and the feature map has a size of 80x80x256, and the feature map has a size of 40x40x256, and the feature map has a size of 40x40x768, and the feature map has a size of 40x40x512, and the feature map has a size of 40x40x512, and the feature map has a size of 20x20x512, and the feature map has a size of 20x20x1024, and the feature map has a size of 20x20x512.

[0026] Refer to Figure 3, both the first DC2f layer and the second DC2f layer include a first DCBS layer, a Split layer, multiple Bottleneck layers, a first concat layer, and a second DCBS layer connected in sequence. The Split layer is connected to each Bottleneck layer respectively.

[0027] Based on the YOLOv8l model parameters and taking the input image size of 640×640 as an example, the input of the 4th layer is 80x80x256, n = 6*d = 6*1 = 6. This layer performs the following operations in sequence: Do a DConv once, and the output is 80x80x256. Based on the number of channels 256, it is divided into two parts, and the y array is y0, y1, both with the dimension of 80x80x128. Do a Bottleneck on y1 (the number of input and output channels remains unchanged), and its output y2 is appended to the array. Then the y array becomes y0, y1, y2. Do six Bottlenecks on y2 again (because n = 6), and the y array becomes y0, y1, y2, y3, y4, y5, y6, y7, all with the dimension of 80x80x128. Connect the y array based on the channel dimension to become 80x80x1028, where 1028 = 128x(n + 2). Do a DConv on the above-connected output to reduce the dimension to 80x80x256.

[0028] See Figure 4 , both the first DCBS layer and the second DCBS layer include a deformable convolutional layer, a BN layer, and a SiLU layer connected in sequence.

[0029] See Figure 5 , the SPPF-SimAM layer includes a CBS layer, multiple MaxPool layers, a concat layer, a Cbs layer, and a SIMAM layer connected in sequence. Among them, the first CBS layer is also connected to each MaxPool layer respectively. The SPPF-SimAM structure can dynamically adjust the weight of each pixel by calculating the similarity between each pixel in the feature map and its surrounding pixels. This calculation method of local self-similarity enables the model to focus on the key regions in the image and enhance the feature representation of these regions. And since the SIMAM layer is a parameter-free attention mechanism, it does not introduce additional computational burden and can improve the ability of the container damage inspection model to capture image details, which is particularly important for tasks such as damage inspection object detection and image segmentation that require fine feature representation.

[0030] In the description of the present application, the descriptions referring to terms such as "one embodiment", "some embodiments", "in this embodiment", "specific example", or "some examples", etc., mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0031] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for automatically checking container damage based on pattern recognition, characterized in that: The following steps are involved: Step S1, for any container body, obtaining a body image of each surface of the container body, preprocessing each body image to obtain a corresponding preprocessed body image, and dividing the preprocessed body image into a training set and a test set; Step S2, using a deep learning framework to build a container damage inspection network, the container damage inspection network includes a backbone module, a feature fusion module and a damage recognition module, the backbone module receives the input image and outputs the feature map , feature map and feature map , the feature map is fused by the feature fusion module , the feature map and the characteristic graph Perform pixel weight adjustment, channel concatenation, and convolution operations to obtain feature maps , feature map and feature map , the feature map is identified by the damage recognition module , the feature map and the characteristic graph Processing is performed to obtain a container damage recognition image; Step S3, using the training set to train the container damage inspection network to obtain the container damage recognition image corresponding to each of the preprocessed container images in the training set; Step S4, training the container damage inspection network for multiple rounds according to step S3 to obtain a container damage inspection network model; Step S5, testing each of the preprocessed container images in the test set through the container damage inspection network model, and outputting the container damage identification image corresponding to each of the preprocessed container images as the automatic container damage inspection result.

2. The method for automatically checking container damage according to claim 1, characterized in that: In step S1, the process of preprocessing each of the box images includes: Each of the box images is mirror-flipped, rotated, scaled, or translated to obtain a plurality of pre-processed box images.

3. The method for automatically checking container damage according to claim 1, characterized in that: The backbone module includes a first CBS layer, a second CBS layer, a first C2f layer, a third CBS layer, a second C2f layer, a fourth CBS layer, a first DC2f layer, a fifth CBS layer and a second DC2f layer connected in sequence, and receives the input image through the first CBS layer and outputs a feature map , receiving the feature map through the second CBS layer And output feature map , receiving the feature map through the first C2f layer And output feature map , receiving the feature map through the third CBS layer And output feature map , receiving the feature map through the second C2f layer And output feature maps respectively and the characteristic graph , receiving the feature map through the fourth CBS layer And output feature map , receiving the feature map through the first DC2f layer And output feature maps respectively and the characteristic graph , receiving the feature map through the fifth CBS layer And output feature map , receiving the feature map through the second DC2f layer And output the feature map , the size of the input image is 640x640x3, the feature map The size of the feature map is 320x320x64. The size of the feature map is 160x160x128. The size of the feature map is 160x160x128. The size of the feature map is 80x80x256. The size of the feature map is 80x80x256. The size of the feature map is 80x80x256. The size of the feature map is 40x40x512. The size of the feature map is 40x40x512. The size of the feature map is 40x40x512. The size of the feature map is 20x20x512. The dimensions are 20x20x512.

4. The method for automatically checking container damage according to claim 3, characterized in that: The first DC2f layer and the second DC2f layer each include a first DCBS layer, a Split layer, a plurality of Bottleneck layers, a first concat layer and a second DCBS layer which are connected in sequence, and the Split layer is connected to each of the Bottleneck layers respectively.

5. The method for automatically checking container damage according to claim 3, characterized in that: The first DC2f layer and the second DC2f layer each include a first DCBS layer, a Split layer, a plurality of Bottleneck layers, a first concat layer and a second DCBS layer which are connected in sequence, and the Split layer is connected to each of the Bottleneck layers respectively.

6. The method for automatically checking container damage according to claim 4, characterized in that: Each of the Bottleneck layers includes two CBS layers and an add layer connected in sequence, and the add layer is connected to the first CSB layer.

7. The method for automatically checking container damage according to claim 1, characterized in that: The feature fusion module includes a SPPF-SimAM layer, a first CARAFE layer, a second concat layer, a third C2f layer, a second CARAFE layer, a third concat layer, a fourth C2f layer, a sixth CBS layer, a fourth concat layer, a fifth C2f layer, a seventh CBS layer, a fifth concat layer and a sixth C2f layer connected in sequence, the SPPF-SimAM layer is also connected to the fifth concat layer, the third C2f layer is also connected to the fourth concat layer, and the feature map is received through the SPPF-SimAM layer. And output feature maps respectively and feature map , receiving the feature map through the first CARAFE layer And output feature map , receiving the feature map through the second concat layer and the characteristic graph And output feature map , receiving the feature map through the third C2f layer And output feature maps respectively and feature map , receiving the characteristic map through the second CARAFE layer And output feature map , receiving the feature map through the third concat layer and the characteristic graph And output feature map , receiving the feature map through the fourth C2f layer And output feature map and the characteristic graph , receiving the feature map through the sixth CBS layer And output feature map , receiving the feature map through the fourth concat layer and the characteristic graph And output feature map , receiving the feature map through the five C2f layers And output feature map and the characteristic graph , receiving the feature map through the seventh CBS layer And output feature map , receiving the feature map through the fifth concat layer and the characteristic graph And output the feature map .

8. The method for automatically checking container damage according to claim 7, characterized in that: The feature map The size of the feature map is 20x20x512. The size of the feature map is 20x20x512. The size of the feature map is 40x40x512. The size of the feature map is 40x40x1024. The size of the feature map is 40x40x512. The size of the feature map is 40x40x512. The size of the feature map is 80x80x512. The size of the feature map is 80x80x768. The size of the feature map is 80x80x256. The size of the feature map is 80x80x256. The size of the feature map is 40x40x256. The size of the feature map is 40x40x768. The size of the feature map is 40x40x512. The size of the feature map is 40x40x512. The size of the feature map is 20x20x512. The size of the feature map is 20x20x1024. The dimensions are 20x20x512.

9. The method for automatically checking container damage according to claim 1, characterized in that: The damage recognition module includes three Detect layers, through which the feature maps are received respectively. , the feature map , the feature map Defects of the prediction frame are identified and marked to obtain the container damage identification image.

10. The method for automatically checking container damage according to claim 9, characterized in that: Each of the Detect layers includes two branches, and each branch only includes two CBS layers, a Conv layer and a Bbox Loss layer connected in sequence.

Images