Scrap monomer example segmentation method and scrap intelligent grading system

Abstract

The application provides a waste steel monomer instance segmentation method and a waste steel intelligent grading system. The waste steel monomer instance segmentation method comprises the following steps: extracting image features in a waste steel image; generating a target candidate box of each image feature; generating a grade of a monomer, a material type classification of the monomer, a detection box of the monomer and a monomer region mask based on each target candidate box. Based on the above detection result, each target detection box is converted into a boundary line of an image monomer, and a mask region of the monomer is generated, so as to complete the monomer instance segmentation. Through the pre-storage of the material type information and the grade information, the subsequent material type and grade classification of the waste steel monomer can be facilitated according to the pre-stored material type information and grade information, so that the waste steel grading is more accurate, and the accuracy of the waste steel classification is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence computer vision technology, and more specifically, to a method for segmenting individual scrap steel instances and a scrap steel intelligent grading system. Background Technology

[0002] Steel companies use various types of scrap steel provided by suppliers for steelmaking. Different grades of scrap steel correspond to different prices. Traditional scrap steel grading methods rely on manual visual judgment, which is highly subjective and has low accuracy. Intelligent scrap steel grading systems replace traditional manual grading with artificial intelligence, offering faster speed and more objective results. In these systems, the grade of a whole vehicle is determined by instance segmentation of a focused image. However, current intelligent scrap steel grading technologies suffer from significant differences in the classification of individual instances, resulting in low accuracy.

[0003] Therefore, how to propose a solution to improve the accuracy of grade classification for intelligent grading of individual scrap steel instances has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, the first aspect of this invention proposes a method for segmenting individual scrap steel instances.

[0005] A second aspect of the invention also proposes a scrap steel individual instance segmentation system.

[0006] A third aspect of the present invention also proposes a scrap steel individual instance segmentation system.

[0007] A fourth aspect of the invention also proposes a readable storage medium.

[0008] The fifth aspect of the present invention also proposes an intelligent scrap steel grading system.

[0009] In view of this, the first aspect of the present invention proposes a method for segmenting individual scrap steel instances, comprising: extracting image features from a scrap steel image and generating target candidate boxes based on the image features; generating material type information, level information, detection box, and region mask of the individual target based on the target candidate box features; visualizing the material type and level information of the individual target based on the material type and level information, detection box, and region mask, and generating the boundary line and region range of the individual target, thereby completing the segmentation of individual instances.

[0010] According to the scrap steel unit instance segmentation method provided by the present invention, image features are extracted from the scrap steel image to identify the scrap steel units in the image, thereby generating target candidate boxes for each image feature as preliminary border lines for subsequent precise selection of image features. Then, pre-selected and stored material type and grade information are used to classify the material type and grade of each image feature. For example, it distinguishes which material type the image feature belongs to, such as spring plate, engine, front and rear axle, etc., and which grade the corresponding scrap steel unit belongs to, such as Special Grade 1, Special Grade 2, Heavy Grade 1, Heavy Grade 2, Heavy Grade 3, Medium Grade, etc., thereby completing the classification of each unit's material type and grade. Based on the material type and grade classification results of the image features, border lines are generated for each image feature on the scrap steel image. The detection box generation visualizes the target candidate box as a border line, which precisely selects the image feature. Simultaneously, a mask is generated based on the target candidate box features and overlaid within the border line to effectively distinguish the image feature. This application enables the pre-storage of material type and grade information, facilitating subsequent classification of individual scrap steel units based on the pre-stored material type and grade information. This makes scrap steel grading more accurate and improves the accuracy of scrap steel classification.

[0011] In addition, the scrap steel single-item segmentation method in the above-mentioned technical solution provided by the present invention may also have the following additional technical features:

[0012] In the above technical solution, the material type information includes: feature information of the corresponding layer of image features, and target candidate boxes of the corresponding layer are generated based on FPN.

[0013] In this technical solution, feature layers corresponding to each image feature are generated, so that each image feature has its own feature layer. This facilitates subsequent information annotation on different feature layers and avoids data confusion caused by multiple feature annotations on the same feature layer. At the same time, target candidate boxes are generated on different feature layers based on FPN. The target candidate boxes are of multiple scales and can be applied to image features of various sizes. This facilitates subsequent transformation of the target candidate boxes so that the eigenvalues ​​corresponding to the feature layer can be accurately selected.

[0014] In the above technical solution, the level information includes: generating target candidate boxes based on FPN; and generating level information by fusing the features of the target candidate boxes with material type information.

[0015] In this technical solution, the target candidate box and material type information generated by FPN can generate corresponding level information, and then perform level analysis on different monomers to determine the level information corresponding to different monomers.

[0016] In the above technical solution, generating target candidate boxes based on FPN for image features includes: setting a preset number of anchors at any point in the feature layer, with each anchor having multiple different aspect ratios and receptive field sizes, thereby generating target candidate boxes. In this technical solution, by setting conditions for generating target candidate boxes before generating each image feature's target candidate box, i.e., preset information, the network can generate target candidate boxes based on the preset information. Specifically, a preset number of anchors are set at each point in each feature layer according to actual needs, and each anchor has a preset number of different aspect ratios and different receptive field sizes. This ensures that the network has good detection capabilities for detection targets of different sizes, i.e., image features, thereby generating the target candidate boxes.

[0017] In the above technical solution, the material type information and level information of the single target are generated based on the features of the target candidate box, including: material type classification and level classification of the target single entity in the image based on cross-entropy loss and focal loss function.

[0018] In this technical solution, when classifying image features by material type and level, cross-entropy loss and focal loss functions are used for classification. These two functions have the function of measuring the difference between two probability distributions, which can effectively solve the problem of uneven sample distribution.

[0019] In the above technical solution, generating detection boxes for target entities in the corresponding image based on target candidate box features includes: optimizing target detection capability based on CIOU loss function to generate detection boxes for entities with high confidence.

[0020] In this technical solution, the CIOU loss function is used to transform the target candidate box, thereby converting the target candidate box into a bounding line, so as to accurately select image features and improve the accuracy of target single-unit box selection.

[0021] In the above technical solution, generating a mask based on the target candidate box features includes: generating a mask based on image features using a binary cross-entropy loss function.

[0022] In this technical solution, a mask for image features is generated by using a binary cross-entropy loss function to cover the bounding lines of the image features, thereby facilitating the differentiation of different image features.

[0023] A second aspect of the present invention provides a scrap steel single-unit instance segmentation system, comprising: a feature extraction module for extracting image features from a scrap steel image; a candidate box generation module for generating candidate boxes for each target; a material type information generation module for generating material type information of the target single unit in the image; a grade information generation module for fusing material type information and candidate box information to generate grade information of the target single unit in the image; a detection and segmentation module for generating detection boxes and region masks for the targets; and a visualization module for visualizing the above-mentioned material type and grade detection boxes and region masks, thereby completing single-unit instance segmentation.

[0024] The scrap steel single instance segmentation system provided by the technical solution of the present invention includes a feature extraction module, a candidate box generation module, a material type information generation module, a grade information generation module, a detection and segmentation module, and a visualization module. The feature extraction module is used to extract image features from the scrap steel image. The candidate box generation module is used to generate candidate boxes for each target. The material type information generation module is used to generate material type information for each single instance in the image. The grade information generation module is used to fuse material type information and candidate box information to generate grade information for the target single instance in the image. The detection and segmentation module is used to generate detection boxes and region masks for the targets. The visualization module is used to visualize the above-mentioned material type and grade detection boxes and region masks, thereby completing the single instance segmentation. Furthermore, the scrap steel single instance segmentation system provided by the technical solution of the present invention, since it implements the steps of the scrap steel single instance segmentation method provided in the first aspect of the present invention, possesses all the technical effects of the scrap steel single instance segmentation method, and will not be elaborated further here.

[0025] In the above technical solution, the candidate box generation module is also used to generate candidate boxes for each target.

[0026] In this technical solution, feature layers corresponding to each image feature are generated, so that each image feature has its own feature layer. This facilitates subsequent information annotation on different feature layers and avoids data confusion caused by multiple feature annotations on the same feature layer. At the same time, target candidate boxes are generated on different feature layers based on FPN. The target candidate boxes are of multiple scales and can be applied to image features of various sizes. This facilitates subsequent transformation of the target candidate boxes so that the eigenvalues ​​corresponding to the feature layer can be accurately selected.

[0027] In the above technical solution, the candidate box generation module is specifically used to: generate target candidate boxes for each image feature based on preset information; wherein, the preset information includes setting a preset number of anchors at each point of each feature layer, and setting a preset number of different aspect ratios and receptive field sizes for each anchor.

[0028] In this technical solution, the conditions for generating target candidate boxes are set before generating target candidate boxes for each image feature, i.e., preset information, so that target candidate boxes can be generated according to the preset information. Specifically, a preset number of anchors are set at each point of each feature layer according to actual needs, and each anchor is set with a preset number of different aspect ratios and different receptive field sizes, so that the network has good detection capabilities for detection targets of different sizes, i.e., image features.

[0029] In the above technical solution, the grade information generation module is also used for: material type classification and grade classification based on image features using cross-entropy loss and focal loss functions.

[0030] In this technical solution, when classifying image features by material type and level, cross-entropy loss and focal loss functions are used for classification. These two functions have the function of measuring the difference between two probability distributions, which can effectively solve the problem of uneven sample distribution and thus improve the accuracy of level determination.

[0031] In the above technical solution, the detection and segmentation module is also used to: convert each target candidate box into a bounding box of image features based on the CIOU loss function.

[0032] In this technical solution, the CIOU loss function is used to transform the target candidate box, thereby converting the target candidate box into a bounding line, so as to accurately select image features and improve the accuracy of image feature selection.

[0033] In the above technical solution, the detection and segmentation module is also used to generate a mask of image features based on the binary cross-entropy loss function, thereby improving the accuracy of image feature selection.

[0034] In this technical solution, a mask for image features is generated by using a binary cross-entropy loss function to cover the bounding lines of the image features, thereby facilitating the differentiation of different image features.

[0035] A third aspect of the present invention provides a scrap steel single instance segmentation system, comprising: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps defined by the scrap steel single instance segmentation method of any of the above technical solutions.

[0036] The scrap steel unit instance segmentation system provided by the technical solution of the present invention includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps defined in any of the aforementioned scrap steel unit instance segmentation methods. Furthermore, since the scrap steel unit instance segmentation system of this application can implement the steps defined in any of the aforementioned scrap steel unit instance segmentation methods, the scrap steel unit instance segmentation system provided by this technical solution possesses all the beneficial effects of the scrap steel unit instance segmentation methods provided in any of the aforementioned technical solutions.

[0037] A fourth aspect of the present invention provides a readable storage medium having a program and / or instructions stored thereon, wherein the program and / or instructions, when executed by a processor, implement the steps of the scrap steel individual instance segmentation method in any of the above technical solutions.

[0038] The readable storage medium provided by the technical solution of the present invention has all the beneficial technical effects of the above-mentioned scrap steel single instance segmentation method when the program and / or instructions stored thereon are executed by the processor. Therefore, it can implement the steps of the scrap steel single instance segmentation method described above.

[0039] The fifth aspect of the present invention provides a scrap steel intelligent grading system, including a scrap steel individual instance segmentation system as described in the above technical solutions; or a readable storage medium as described in the above technical solutions.

[0040] The intelligent scrap steel grading system provided by the present invention includes either a scrap steel individual instance segmentation system or a readable storage medium as described in the above-described technical solution. Therefore, the intelligent scrap steel grading system possesses all the technical effects of the scrap steel individual instance segmentation system or the readable storage medium, which will not be elaborated further here.

[0041] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0042] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0043] Figure 1 This is a flowchart illustrating a method for segmenting individual scrap steel instances according to an embodiment of the present invention;

[0044] Figure 2 This is a block diagram of a scrap steel individual instance segmentation system according to an embodiment of the present invention;

[0045] Figure 3 This is a block diagram of a scrap steel individual instance segmentation system according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of scrap steel images according to another embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of single-unit instance segmentation according to another embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of a single scrap steel unit according to another embodiment of the present invention;

[0049] Figure 7 This is an overall framework diagram of a scrap steel individual instance segmentation method according to another embodiment of the present invention;

[0050] Figure 8 This is a schematic diagram of the Backbone module network structure according to another embodiment of the present invention;

[0051] Figure 9 This is a schematic diagram of anchor settings according to another embodiment of the present invention;

[0052] Figure 10 This is a schematic diagram of an attention mechanism according to another embodiment of the present invention.

[0053] in, Figure 2 and Figure 3 The correspondence between the reference numerals and component names in the attached drawings is as follows:

[0054] 200 Scrap Steel Individual Instance Segmentation System, 202 Feature Extraction Module, 204 Candidate Box Generation Module, 206 Material Type Information Generation Module, 207 Grade Information Generation Module, 208 Detection and Segmentation Module, 209 Visualization Module, 300 Scrap Steel Individual Instance Segmentation System, 302 Memory, 304 Processor. Detailed Implementation

[0055] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0057] The following reference Figures 1 to 10 A scrap steel intelligent grading system describes a method for segmenting individual scrap steel instances in some embodiments of the present invention.

[0058] The first aspect of this invention proposes a method for segmenting individual scrap steel instances, such as... Figure 1 As shown, it includes:

[0059] S102, Extract image features from scrap steel images;

[0060] S104, Generate target candidate boxes for the image;

[0061] S106, Generate the material type and level information of the monomer based on the target candidate box features, the detection box of the monomer target and the region mask;

[0062] S108: Based on the material type and level information of the individual target, the detection box of the individual target, and the material type and level information of the individual target visualized by the region mask, the bounding line and region range of the individual target are generated to complete the segmentation of the individual instance.

[0063] According to the scrap steel unit instance segmentation method provided in this embodiment, image features are extracted from the scrap steel image to identify the scrap steel units in the image, thereby generating target candidate boxes for each image feature as preliminary border lines for subsequent precise image feature selection. Then, the material type and level of each image feature are classified by pre-selected and stored material type and level information. For example, it is distinguished which material type the image feature belongs to, such as spring plate, engine, front and rear axle, etc., and which level the scrap steel unit corresponding to the image feature belongs to, such as special grade 1, special grade 2, heavy grade 1, heavy grade 2, heavy grade 3, medium grade, etc., thereby completing the classification of the material type and level of each unit. Then, based on the material type and level classification results of the image features, border lines are generated for each image feature on the scrap steel image. The detection box generation visualizes the target candidate box as a border line, and the border line accurately selects the image feature. At the same time, a mask is generated based on the target candidate box features and covered within the border line to enable good differentiation of the image feature. This application enables the pre-storage of material type and grade information, facilitating subsequent classification of individual scrap steel units based on the pre-stored material type and grade information. This makes scrap steel grading more accurate and improves the accuracy of scrap steel classification.

[0064] In the above embodiments, after the step of extracting image features from the scrap steel image, the method further includes: generating feature layers corresponding to each image feature; and generating target candidate boxes for each feature layer based on FPN.

[0065] In this embodiment, by generating feature layers corresponding to each image feature, each image feature has its own feature layer, which facilitates subsequent information annotation on different feature layers and avoids data confusion caused by multiple feature annotations on the same feature layer. At the same time, target candidate boxes are generated on different feature layers based on FPN. The target candidate boxes are of multiple scales and can be applied to image features of various sizes, so as to facilitate subsequent transformation of the target candidate boxes and accurately select the eigenvalues ​​corresponding to the feature layer.

[0066] In the above embodiments, the level information includes: generating target candidate boxes based on FPN; and generating level information by fusing the features of the target candidate boxes with material type information.

[0067] In this embodiment, the target candidate box and material type information generated by FPN can generate corresponding level information, and then perform level analysis on different monomers to determine the level information corresponding to different monomers.

[0068] In the above embodiments, the step of generating target candidate boxes for each image feature specifically includes: generating target candidate boxes for each image feature based on preset information; wherein, the preset information includes setting a preset number of anchors at each point of each feature layer, and setting a preset number of different aspect ratios and receptive field sizes for each anchor.

[0069] In this embodiment, before generating target candidate boxes for each image feature, the conditions for generating target candidate boxes are set, i.e., preset information, so that target candidate boxes can be generated according to the preset information. Specifically, a preset number of anchors are set at each point of each feature layer according to actual needs, and each anchor is set with a preset number of different aspect ratios and different receptive field sizes, so that the network has good detection capabilities for detection targets of different sizes, i.e., target units in the image.

[0070] In the above embodiments, the scrap steel individual instance segmentation method further includes: classifying the material type and level of the target individual in the image based on cross-entropy loss and focal loss functions.

[0071] In this embodiment, when classifying image features by material type and level, cross-entropy loss and focal loss functions are used for classification. These two functions have the function of measuring the difference between two probability distributions, which can effectively solve the problem of uneven sample distribution.

[0072] In the above embodiments, the scrap steel individual instance segmentation method further includes: converting each target candidate box into a bounding box line of the target individual in the image based on the CIOU loss function.

[0073] In this embodiment, the CIOU loss function is used to transform the target candidate boxes, thereby converting the target candidate boxes into border lines to accurately select target units in the image and improve the accuracy of target unit selection in the image.

[0074] In the above embodiments, the scrap steel individual instance segmentation method further includes: generating an image feature mask based on a binary cross-entropy loss function.

[0075] In this embodiment, a mask for image features is generated by using a binary cross-entropy loss function to cover the bounding lines of the image features, thereby facilitating the differentiation of different image features.

[0076] A second aspect of the present invention provides a scrap steel individual instance segmentation system 200, such as... Figure 2 As shown, the system includes: a feature extraction module 202, a candidate box generation module 204, a material type information generation module 206, a grade information generation module 207, a detection and segmentation module 208, and a visualization module 209. Specifically, the feature extraction module 202 extracts image features from the scrap steel image. The candidate box generation module 204 generates candidate boxes for each target. The material type information generation module 206 generates material type information for each individual unit in the image. The grade information generation module 207 fuses the material type information and candidate box information to generate grade information for the target units in the image. The detection and segmentation module 208 generates detection boxes and region masks for the targets. The visualization module 209 visualizes the aforementioned material type and grade detection boxes and region masks, thereby completing the segmentation of individual instances.

[0077] The scrap steel single instance segmentation system 200 provided according to an embodiment of the present invention includes a feature extraction module 202, a candidate box generation module 204, a material type information generation module 206, a grade information generation module 207, a detection and segmentation module 208, and a visualization module 209. The feature extraction module 202 is used to extract image features from the scrap steel image. The candidate box generation module 204 is used to generate candidate boxes for each target. The material type information generation module 206 is used to generate material type information for each single instance in the image. The grade information generation module 207 is used to fuse material type information and candidate box information to generate grade information for the target single instance in the image. The detection and segmentation module 208 is used to generate detection boxes and region masks for the targets. The visualization module 209 is used to visualize the above-mentioned material type and grade detection boxes and region masks, thereby completing the single instance segmentation. Furthermore, the scrap steel single instance segmentation system provided according to the embodiment of the present invention, since it implements the steps of the scrap steel single instance segmentation method provided in the first aspect of the present invention, possesses all the technical effects of the scrap steel single instance segmentation method, and will not be elaborated further here.

[0078] In the above embodiments, the material pattern information generation module 206 is further configured to: generate feature layers corresponding to each image feature; and generate target candidate boxes for each feature layer based on FPN.

[0079] In this embodiment, by generating feature layers corresponding to each image feature, each image feature has its own feature layer, which facilitates subsequent information annotation on different feature layers and avoids data confusion caused by multiple feature annotations on the same feature layer. At the same time, target candidate boxes are generated on different feature layers based on FPN. The target candidate boxes are of multiple scales and can be applied to image features of various sizes, so as to facilitate subsequent transformation of the target candidate boxes and accurately select the eigenvalues ​​corresponding to the feature layer.

[0080] In the above embodiments, the candidate box generation module 204 is specifically used to: generate target candidate boxes for each image feature based on preset information; wherein, the preset information includes setting a preset number of anchors at each point of each feature layer, and setting a preset number of different aspect ratios and receptive field sizes for each anchor.

[0081] In this embodiment, before generating the target candidate box for each image feature, the conditions for generating the target candidate box are set, i.e., preset information, so that the target candidate box can be generated according to the preset information. Specifically, a preset number of anchors are set at each point of each feature layer according to actual needs, and each anchor is set with a preset number of different aspect ratios and different receptive field sizes, so that the network has good detection capabilities for detection targets of different sizes, i.e., image features.

[0082] In the above embodiments, the grade information generation module 207 is also used for: performing material type classification and grade classification based on image features using cross-entropy loss and focal loss functions.

[0083] In this embodiment, when classifying image features by material type and level, cross-entropy loss and focal loss functions are used for classification. These two functions have the function of measuring the difference between two probability distributions, which can effectively solve the problem of uneven sample distribution.

[0084] In the above embodiments, the detection and segmentation module 208 is further configured to: convert each target candidate box into a bounding box of image features based on the CIOU loss function.

[0085] In this embodiment, the CIOU loss function is used to transform the target candidate box, thereby converting the target candidate box into a bounding line, so as to accurately select image features and improve the accuracy of image feature selection.

[0086] In the above embodiments, the detection and segmentation module 208 is also used to: generate a mask of image features based on the binary cross-entropy loss function, thereby improving the accuracy of the determination level.

[0087] In this embodiment, a mask for image features is generated by using a binary cross-entropy loss function to cover the bounding lines of the image features, thereby facilitating the differentiation of different image features.

[0088] A third aspect of the present invention provides a scrap steel individual instance segmentation system 300, such as... Figure 3 As shown, it includes: a memory 302, a processor 304, and a program stored in the memory 302 and executable on the processor 304. When the program is executed by the processor 304, it implements the steps defined by the scrap steel single-item segmentation method of any of the above embodiments.

[0089] The scrap steel unit instance segmentation system provided by the embodiments of the present invention includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps defined in any of the above-described scrap steel unit instance segmentation methods. Furthermore, since the scrap steel unit instance segmentation system of this application can implement the steps defined in any of the above-described scrap steel unit instance segmentation methods, the scrap steel unit instance segmentation system provided in this embodiment has all the beneficial effects of the scrap steel unit instance segmentation methods provided in any of the above embodiments.

[0090] A fourth aspect of the present invention provides a readable storage medium having a program and / or instructions stored thereon, which, when executed by a processor, implement the steps of the scrap steel individual instance segmentation method in any of the above embodiments.

[0091] The readable storage medium provided according to the embodiments of the present invention has all the beneficial technical effects of the above-described scrap steel single instance segmentation method when the program and / or instructions stored thereon are executed by a processor. Therefore, it will not be repeated here.

[0092] A fifth aspect of the present invention provides a scrap steel intelligent grading system, including a scrap steel individual instance segmentation system as described in the above embodiments; or a readable storage medium as described in the above embodiments.

[0093] The intelligent scrap steel grading system provided according to embodiments of the present invention includes a scrap steel individual instance segmentation system as described in the above embodiments or a readable storage medium as described in the above embodiments. Therefore, the intelligent scrap steel grading system possesses all the technical effects of the scrap steel individual instance segmentation system or the readable storage medium, which will not be elaborated further here.

[0094] The following describes the scrap steel single-unit instance segmentation method provided in this application in conjunction with another specific embodiment.

[0095] This embodiment proposes a method for segmenting individual scrap steel instances based on prior information, which achieves accurate classification of scrap steel grades by introducing prior material type information of individual scrap steel instances.

[0096] The intelligent scrap steel grading system mainly includes the following modules:

[0097] (1) Automatic Focusing Carriage Photography Module. A PTZ camera is installed at the unloading point to focus and photograph the scrap steel in the carriage, obtaining an 800w pixel image of the scrap steel, such as... Figure 4 As shown.

[0098] (2) Scrap Steel Individual Segmentation Module. Using instance segmentation algorithms from the field of computer vision, each individual scrap steel in the image is classified, detected, and segmented to obtain the level (S, A, 0, 1, 2, 3, 4, 5), material type (steel plate, pipe, cut parts, etc.), location, and region of each individual scrap steel. Figure 5 As shown.

[0099] (3) Density parameter module. The density of a certain material type is measured on-site, and then the area and thickness of the monomer are calculated based on the results of (2). Finally, the weight of the monomer is obtained, and the weight ratio of the whole vehicle is obtained after comprehensive analysis.

[0100] This invention pertains to the individual scrap steel segmentation module within an intelligent scrap steel grading system.

[0101] The classification branch in the scrap steel unit segmentation module determines the grade of each scrap steel unit, such as Special Grade 1, Special Grade 2, Heavy Grade 1, Heavy Grade 2, Heavy Grade 3, Medium, etc. If only the grade is labeled, units within the same grade will vary greatly (e.g., leaf springs, engine parts, and axles all belong to Special Grade 1), posing significant challenges to classification. An example image of a scrap steel unit is shown below. Figure 6 As shown.

[0102] The present invention aims to propose a method for segmenting scrap steel individual units based on prior information, which makes full use of the material type information of scrap steel individual units as prior knowledge to assist in the determination of the grade of individual units.

[0103] The scrap steel individual instance segmentation method proposed in this invention includes the following steps:

[0104] The Backbone module extracts image features;

[0105] The RPN module uses FPN to generate target candidate boxes at multiple feature layers;

[0106] The Attention module utilizes the features of material shape information and fuses them with level branches through CBAM to assist in level classification;

[0107] The Head module is used to predict the target category, location, and region.

[0108] The overall algorithm framework process, such as Figure 7 As shown.

[0109] The Backbone module consists of an HRNet network structure and is used to extract image features. Its network structure is as follows: Figure 8 As shown. The RPN module sets K anchors at each point in the FPN feature layer, such as... Figure 9 As shown, multiple aspect ratios and receptive field sizes are set for the anchors. By setting anchors on different feature layers, the network can effectively detect targets of both large and small sizes. To improve the accuracy of single-unit level classification, this invention introduces prior information on material type to assist in level classification. The feature layer of each candidate box is divided into two branches: one branch predicts material type classification, and the other branch predicts level classification. The feature layer of the material type branch is fused with the level branch through an Attention module to improve the expressive power of level features. The attention mechanism module is as follows: Figure 10 As shown, the Head module produces the output for the segmentation task. The Head module includes three branches: one for classification, one for predicting bounding boxes, and one for generating individual unit masks. The classification branch is further divided into material type classification and level classification. The objective function of the classification branch is cross-entropy loss and focal loss to address the imbalanced sample distribution problem; the objective function of the detection branch is CIOU loss; and the loss function of the segmentation branch is binary cross-entropy loss.

[0110] In this specification, the term "multiple" refers to two or more unless otherwise expressly defined. The terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0111] In the description of this specification, the terms "one embodiment," "some embodiments," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0112] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A scrap steel monomer instance segmentation method for a scrap steel intelligent grading system, characterized in that, The method comprises the following steps: extracting image features in a scrap steel image; generating a target candidate box of the image features; generating the shape information of a single target, the level information of the single target, the detection box of the single target and the region mask based on the target candidate box features; visualizing the shape and level information of the single target based on the shape and level information of the single target, the detection box of the single target and the region mask, generating the bounding box line and the region range of the single target, and thus completing the single instance segmentation; the generation of the shape information and the level information of the single target based on the target candidate box features comprises: classifying the shape and level of the image features based on the cross-entropy loss and the focal loss function, and then obtaining the shape classification information and the level classification information of the single target; the generation of the detection box of the single target based on the target candidate box features comprises: training the detection target function based on the CIOU loss function to generate the detection box of the single target; the visualization of the shape and level information of the single target based on the shape and level information of the single target, the detection box of the single target and the region mask, and the generation of the bounding box line and the region range of the single target to complete the single instance segmentation comprises: the detection box is generated to visualize the target candidate box as a bounding box line, which accurately selects the image features, and a mask is generated based on the target candidate box features and overlaid in the bounding box line; the generation of the shape information and the level information of the single target based on the target candidate box features comprises: the Attention module fuses the features of the shape information with the level branch through CBAM to assist the level classification.

2. The worn single body instance segmentation method of claim 1, wherein, the generation of the shape information and the level information of the single target based on the target candidate box features comprises: the feature information of the corresponding layer of the image features; generating the target candidate box of the corresponding layer based on FPN.

3. The waste monomer instance segmentation method according to claim 2, wherein, the generation of the shape information and the level information of the single target based on the target candidate box features further comprises: generating the target candidate box based on FPN; fusing the target candidate box features and the shape information to generate the level information.

4. The waste monomer instance segmentation method according to claim 2 or 3, characterized in that, the generation of the target candidate box of the corresponding layer based on FPN comprises: setting a preset number of anchors at any point of the feature layer, the anchors are set to have multiple different width-height ratios and field sizes, and thus the target candidate box is generated.

5. The worn single body instance segmentation method of claim 1, wherein, generating a region mask based on the target candidate box features comprises: training based on the binary cross-entropy loss function to generate the region mask corresponding to the single target.

6. A system for instance segmentation of scrap singletons, the system comprising: The method comprises the following steps: a feature extraction module for extracting image features in a scrap steel image; a candidate box generation module for generating a candidate box for each single target; a shape information generation module for generating the shape information of the single target in the image; a level information generation module for fusing the shape information and the candidate box information to generate the level information of the single target in the image; The detection segmentation module is configured to generate a detection frame and a region mask of the single target; The visualization module is configured to perform image visualization on the level information, the detection frame and the region mask, so as to complete the single instance segmentation; The type information generation module and the level information generation module are specifically configured to: classify the image features based on a cross-entropy loss and a focal loss function to obtain type classification information and level classification information of the single target; The detection segmentation module is specifically configured to: train a detection target function based on a CIOU loss function to generate a corresponding detection frame of the single target; The visualization module is specifically configured to: The detection frame is generated to visualize the target candidate frame as a frame line, and the frame line accurately selects the image features, and a mask is generated according to the target candidate frame features and is overlaid in the frame line; The level information generation module is further configured to: The Attention module fuses the features of the type information through CBAM and a level branch to assist the level classification.

7. A scrap steel individual instance segmentation system, characterized in that, The program is stored in the memory and executable on the processor, and when the program is executed by the processor, the steps defined in the scrap single instance segmentation method in any one of claims 1 to 5 are implemented. The program and / or instructions are stored on the processor, and when the program and / or instructions are executed by the processor, the steps of the scrap single instance segmentation method in any one of claims 1 to 5 are implemented.

8. A readable storage medium, characterized by, The scrap single instance segmentation system according to claim 6 or 7; or 9. A scrap steel intelligent grading system, characterized in that, The readable storage medium according to claim 8. ​

Images