Lightweight surface crack detection method and system for mineral sintering surface

Abstract

The invention discloses a lightweight surface crack detection method and system for a mineral sintering surface, and the method comprises the steps: carrying out the segmentation and crack classification of a collected mineral sintering surface image, and constructing a sintered ore surface defect data set; constructing a large-selection nuclear crack detection model based on the improved YOLOv7, wherein the large-selection nuclear crack detection model comprises a backbone network, an LSKNet module and an SPPCSPC module; the LSKNet module dynamically adjusts a receptive field range and combines depth separable convolution and a spatial selection mechanism to realize global and local feature fusion; training and testing the model by adopting a data set; and carrying out real-time defect detection on a to-be-detected mineral sintering surface image by utilizing the trained model, outputting the grade of each crack, and calculating the score of the overall severity of the surface based on the grade of each crack. According to the method, the mineral sintering surface defects can be accurately and rapidly detected, reliable data support is provided for intelligent smelting, the production cost is reduced, and the production efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence, object detection, and mineral sintering, and more specifically, to a lightweight surface crack detection method and system for mineral sintered surfaces. Background Art

[0002] As a typical process industry in traditional manufacturing, metal smelting has made significant contributions to infrastructure development and economic growth. To achieve the goals of high-quality, high-yield, and low-consumption steel manufacturing, ore sintering provides raw materials for subsequent purification processes and is a crucial step in metal smelting. However, defects in sintered products can seriously affect the blast furnace smelting process, such as causing unstable working conditions and difficulty in meeting the quality requirements of smelted metals and other by-products. In the future, intelligent sintering processes and measurements are expected to become important research areas in academia and industry, aiming to improve quality, increase productivity, save energy, protect the environment, and achieve sustainable development. The increasing maturity and application of technologies such as artificial intelligence, cyber-physical systems, big data, and cloud computing, as well as their integration with the industrial Internet industrial chain, will surely promote the upgrading of sintering technology, and intelligent sintering systems will become the focus of future development. An important part of an intelligent sintering system is the identification of sintered surface defects. However, surface defects such as cracks, pores, or uneven sintering inevitably occur during the sintering process, and these defects can significantly affect the stability of subsequent smelting and the quality of finished products. For example, defects may cause fluctuations in blast furnace operating conditions, thereby reducing production efficiency and increasing costs.

[0003] Currently, the defect detection of mineral sintered surfaces mainly relies on manual inspection or traditional machine vision-based methods. Manual inspection is highly subjective and inefficient, making it difficult to meet the requirements of modern industry for high-efficiency and accurate detection. Traditional machine vision methods mainly rely on technologies such as edge detection, threshold segmentation, and feature extraction, and are difficult to handle complex sintered surface features and images with high noise and low contrast. In addition, these methods have poor robustness to environmental light changes and interference objects and cannot meet the requirements of actual industrial scenarios.

[0004] Nowadays, accurately identifying defects such as cracks in the mineral sintering surface and detecting defects through object detection are important steps in improving product quality in the intelligent sintering system. In recent years, with the rapid development of artificial intelligence and deep learning technologies, object detection methods based on deep learning have made remarkable progress in multiple fields. Such methods can automatically extract multi-level feature information and have higher detection accuracy and robustness for complex surface defects. However, most of the existing deep learning models are designed for general object detection, with high computational complexity, large model volume, and slow operation speed, which are not suitable for deployment on industrial field devices or mobile terminals with limited resources. In addition, the existing models often do not fully combine the unique characteristics of the mineral sintering surface, resulting in room for improvement in detection accuracy and efficiency. Moreover, due to the shooting equipment, sintering images may have a lot of noise, making it impossible to accurately perform object detection.

[0005] In summary, the existing methods for object detection on the sintering material surface have one or more of the following problems:

[0006] 1. The model complexity is too high, occupying a large amount of computing and storage resources, and cannot be configured on small and medium-sized devices and mobile terminals.

[0007] 2. Due to the particularity of the mineral sintering image surface, there is a lot of noise, the texture features are not processed carefully enough, the overall receptive field is not large enough, and the detection accuracy for cracks is not high. Summary of the Invention

[0008] In view of this, the present invention provides a lightweight surface crack detection method and system for the mineral sintering surface to solve the problems of low detection accuracy, high model complexity, and poor adaptability in the existing detection methods.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] In the first aspect, an embodiment of the present invention provides a lightweight surface crack detection method for the mineral sintering surface, including the following steps:

[0011] S10. Segment and classify cracks in the collected mineral sintering surface image to construct a sintered ore surface defect data set; divide the sintered ore surface defect data set into a training set, a validation set, and a test set;

[0012] S20. Construct a large selection kernel crack detection model based on the improved YOLOv7. The model includes:

[0013] (a) A backbone network, composed of multiple CBS modules, CBM modules, and ELAN modules, for multi-scale feature extraction;

[0014] (b) The LSKNet module realizes the fusion of global and local features by dynamically adjusting the receptive field range, combining depthwise separable convolution and spatial selection mechanism;

[0015] (c) The SPPCSPC module generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintered surface;

[0016] S30. Train the model using the training set labeled with crack classification, optimize the model robustness through data augmentation strategy, adjust the hyperparameters based on the validation set, and evaluate the detection effect based on the test set;

[0017] S40. Use the trained model to perform real-time defect detection on the mineral sintered surface image to be detected, output the level of each crack, and calculate the overall severity score of the surface based on the level of each crack.

[0018] Furthermore, the crack classification in S10 includes:

[0019] It is divided into 1 - 4 categories according to the ratio of crack length to the width of the material surface.

[0020] Furthermore, the CBS module includes a convolutional layer, a batch normalization layer, and a Silu activation function connected in sequence, which is used to balance the fineness of feature crack capture and computational efficiency;

[0021] The CBM module includes a convolutional layer, a batch normalization layer, and a Sigmoid activation function connected in sequence, which is used to predict the confidence of the target crack;

[0022] The ELAN module enhances the expression ability of features at different scales through a multi-path feature fusion mechanism.

[0023] Furthermore, the backbone network also includes:

[0024] Lightweight design, using 1×1 convolutional kernels to replace traditional large convolutional kernels, and reducing redundant layers to reduce computational complexity.

[0025] Furthermore, the LSKNet module includes:

[0026] a) The LSK Selection branch adopts a large convolutional kernel selection mechanism to extract features of the large kernel receptive field through a fully connected layer and a GELU activation function;

[0027] b) The FFN branch combines depthwise separable convolution and GELU activation function to efficiently extract local features.

[0028] Furthermore, the SPPCSPC module includes:

[0029] 1) One 1×1 convolutional layer and three 3×3 dilated convolutional layers for lightweight multi-scale feature extraction;

[0030] 2) One global average pooling layer to obtain image-level features, which are then fed into a 1×1 convolutional layer, and bilinear interpolation is used to obtain an image of the same size;

[0031] 3) The five multi-scale features obtained are merged together through the same channel dimension, and then fed into a 1×1 convolution for post-fusion to obtain feature information with a new number of channels.

[0032] Furthermore, the data augmentation strategy includes random cropping, color transformation, and image rotation to simulate the illumination changes and noise interference in industrial scenarios.

[0033] Furthermore, in step S40, calculating the overall surface severity score based on each crack level includes:

[0034] The formula for the single crack score is as follows:

[0035]

[0036] Among them, the crack level i is divided into levels 1-4 according to the ratio of the crack length to the width of the material surface, corresponding to different reference scores and coefficients;

[0037]

[0038] In the formula, S represents the overall surface defect severity score; S k represents the score of the k-th crack; K represents the total number of cracks in the image.

[0039] In a second aspect, an embodiment of the present invention also provides a lightweight surface crack detection system for the surface of mineral sintering, including:

[0040] A dataset construction module for segmenting and classifying the collected mineral sintering surface images to construct a sintered ore surface defect dataset; dividing the sintered ore surface defect dataset into a training set, a validation set, and a test set;

[0041] A detection module construction module for constructing a large selection kernel crack detection model based on the improved YOLOv7, and the model includes:

[0042] (a) A backbone network, composed of multiple CBS modules, CBM modules, and ELAN modules, for multi-scale feature extraction;

[0043] (b) The LSKNet module, which realizes the fusion of global and local features by dynamically adjusting the receptive field range and combining depthwise separable convolution and spatial selection mechanism;

[0044] (c) SPPCSPC module, which generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintered surface;

[0045] Training and testing module, which is used to train the model with a training set labeled with crack classification, optimize the model robustness through data augmentation strategies, adjust hyperparameters based on the validation set, and evaluate the detection effect based on the test set;

[0046] Detection module, which uses the trained model to perform real-time defect detection on the mineral sintered surface image to be detected, outputs the grade of each crack, and calculates the overall severity score of the surface based on the grade of each crack.

[0047] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following technical advantages:

[0048] 1. Efficient feature extraction: Through the design of a deep neural network, the texture features and defect semantic information of the sintered surface image are fully utilized to improve the accuracy and robustness of defect detection.

[0049] 2. Model lightweight: Aiming at the problem of limited resources of industrial field equipment, the network structure is optimized and designed to reduce the computational amount and storage requirements, and realize the deployment of the model on mobile or small and medium-sized devices.

[0050] 3. Generalization improvement: Design a detection mechanism for the characteristics of the sintered surface to improve the adaptability to complex texture backgrounds, environmental noises and illumination changes.

[0051] 4. Real-time performance: By optimizing the detection process, the detection speed is increased to meet the requirements of real-time detection in the industrial field.

[0052] Through the present invention, accurate and rapid detection of defects on the mineral sintered surface can be realized, providing reliable data support for intelligent smelting, reducing production costs, improving production efficiency, and at the same time promoting the development of the steel industry towards the direction of intelligentization and greenization. Description of the Drawings

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0054] Figure 1 It is a flowchart of a lightweight surface crack detection method for the mineral sintered surface provided by the present invention.

[0055] Figure 2 This is an example diagram for calculating the surface defects of sintered ore provided by the present invention.

[0056] Figure 3 This is the schematic diagram of the large selection kernel crack detection model provided by the present invention.

[0057] Figure 4 This is the structure diagram of the CBS module provided by the present invention.

[0058] Figure 5 This is the structure diagram of the CBM module provided by the present invention.

[0059] Figure 6 This is the structure diagram of the ELAN module provided by the present invention.

[0060] Figure 7 This is the structure diagram of the LSKNet module provided by the present invention.

[0061] Figure 8 This is the structure diagram of the SPPCSPC module provided by the present invention.

[0062] Figure 9 This is the structure diagram of the lightweight surface crack detection system for the mineral sintering surface provided by the present invention. Detailed implementation manners

[0063] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0064] Embodiment:

[0065] Refer to Figure 1 As shown, the embodiment of the present invention discloses a lightweight surface crack detection method for the mineral sintering surface, including:

[0066] S10. Segment and classify the cracks in the collected mineral sintering surface image to construct a sintered ore surface defect data set; divide the sintered ore surface defect data set into a training set, a validation set, and a test set;

[0067] For example, the sintering surface image can be collected by a high-temperature industrial camera module, covering different illuminations, noises, and defect types. The severity level (1-4 levels) is divided according to the ratio of the crack length to the material surface width, and a single crack scoring method is designed. For example, the LabelIMG tool is used to manually label the crack position and level to construct a training set, a validation set, and a test set.

[0068] S20. Construct a large selection kernel crack detection model based on the improved YOLOv7. The model includes:

[0069] (a) A backbone network, composed of multiple CBS modules, CBM modules, and ELAN modules, for multi-scale feature extraction;

[0070] (b) An LSKNet module, which realizes the fusion of global and local features by dynamically adjusting the receptive field range and combining depthwise separable convolution and spatial selection mechanism;

[0071] (c) An SPPCSPC module, which generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintered surface;

[0072] Through dynamically adjusting the receptive field and multi-scale feature fusion, this model can efficiently identify cracks, uneven sintering and other defects on the sintered surface, and has the characteristics of high precision, lightweight, and strong generalization ability.

[0073] S30. Train the model with a training set labeled with crack classification, optimize the model robustness through data augmentation strategies, adjust hyperparameters based on the validation set, and evaluate the detection effect based on the test set. For example, data augmentation uses strategies such as random cropping, color transformation, and rotation to simulate industrial scenario interference. Use CIoU Loss to optimize the boundary box regression accuracy; dynamically adjust parameters such as the learning rate and weight decay based on the validation set to prevent overfitting.

[0074] S40. Use the trained model to perform real-time defect detection on the mineral sintered surface image to be detected, output the level of each crack, and calculate the overall severity score of the surface based on the level of each crack.

[0075] In this step, after inputting the sintered image to be detected, the model extracts multi-scale features through the backbone network, dynamically selects features through LSKNet, and finally the detection head outputs the crack position and crack level, and then the overall severity score of the surface can be further calculated; for example, the detection result can be real-time fed back to the intelligent smelting system to guide the adjustment of process parameters.

[0076] The present invention improves the detection accuracy of cracks on the mineral sintered surface, and can adapt to complex environments; the detection model is lightweight, reducing resource costs; it also enhances the generalization ability and can adapt to various scenarios; it meets the real-time requirements and responds quickly.

[0077] The following details each step of the present invention:

[0078] Step 1: Construct a sintered ore quality evaluation data set;

[0079] To achieve the target detection task of the sintering burden surface and identify cracks of different degrees, it is first necessary to segment and classify the cracks in the collected images, and construct a sintering burden surface dataset based on the surface defect quantitative evaluation model. Among them, referring to the work experience of multiple experts, the surface defect quantitative evaluation model formulates the calculation formula for the score of a single crack and the grading standard for the severity level of a single crack, as shown in Formula (1) and

[0080] Table 1.

[0081]

[0082] Table 1 Grading Standard for the Severity Level of a Single Crack

[0083]

[0084] Due to visual reasons, the obtained sintering image is not rectangular but trapezoidal. Therefore, the front end burden surface of the sinter in the image data can be approximately regarded as an isosceles trapezoid with the upper base of l1, the lower base of l2, and the height of h, that is, as shown in the shaded part in Figure 2 . Among them, l2 is the width of the burden surface image, l1 is the approximate value of the average number of non - black pixels in the burden surface image after semantic segmentation at the uppermost end of the burden surface image, and h is the length of the burden surface image. Based on this assumption, the calculation steps for the severity score S of the sinter burden surface defect are as follows:

[0085] Step 1: For example, use the target detection algorithm to identify all cracks in the burden surface image and obtain the corresponding severity levels of the cracks; in this invention, step S40 can identify the levels of all cracks in the burden surface image. Taking a certain crack A as an example, its corresponding severity level is i ∈ [1, 4], and the coordinates of the detection box are Figure 2 shown as (x1, y1), (x2, y1), (x1, y2), (x2, y2) in sequence.

[0086] Step 2: If multiple cracks are detected in the burden surface image, calculate the areas of all detection boxes identified in the image.

[0087] Step 3: Sort all detection boxes according to the area size. Starting from the detection box with the largest area, analyze its area overlap with the remaining detection boxes in sequence. Denote the areas of the two detection boxes in a certain comparison as S A , S B , and the overlapping area as S AB . If the overlapping area meets the following conditions, then eliminate detection box B, and so on until all detection boxes are traversed.

[0088] S AB ≥0.8S B , S A >S B (2)

[0089] Step 4: Obtain the evaluation scores of all cracks in sequence according to the single crack score calculation formula (1). For example, the score of crack A can be calculated as:

[0090]

[0091] Step 5: If there are K cracks remaining in the burden surface image, the score for the overall surface defect severity is:

[0092]

[0093] Considering the environment at the sintering site, a high-temperature-resistant industrial camera module is required for image acquisition. The complete acquired image is manually marked using LabelIMG, and the annotation objects are the cracks on the sintering surface with different severities. Then, the marked dataset is provided.

[0094] Step 2: Construct a large selection kernel crack detection model for object detection;

[0095] For the object detection task on the sintering surface, the large selection kernel crack detection model is a deep learning model based on the improved YOLOv7 and the large selection kernel network (LSKNet), designed specifically for real-time defect detection in industrial complex scenarios. Through dynamic adjustment of the receptive field and multi-scale feature fusion, this model can efficiently identify defects such as cracks and uneven sintering on the sintering surface, and has the characteristics of high precision, lightweight, and strong generalization ability.

[0096] Refer to Figure 3 As shown, in the feature extraction part of the model, the backbone network adopts multiple improved modules, including the CBS module (convolution, normalization, and Silu activation), the ELAN module (efficient hierarchical feature fusion), and the SPPCSPC module (spatial pyramid pooling). Among them, the CBS module captures both detailed information and global feature modeling through multi-scale convolution operations; the ELAN module enhances the expression ability of shallow and deep features through the feature fusion strategy, providing a rich feature basis for subsequent object detection; while the SPPCSPC module generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintering surface. This modular design not only improves the model's expression ability for complex surface features but also effectively reduces computational redundancy.

[0097] The LSKNet module is one of the core innovations in large selection kernel crack detection. By dynamically adjusting the receptive field range, this module enables the model to better balance the global information of large-scale targets and the detailed information of small-scale targets. In practical applications, for sintered surfaces with widespread cracks, LSKNet can capture their global semantic information; for tiny cracks, it achieves precise detection by focusing on local features. In addition, LSKNet combines channel attention and spatial selection mechanisms, significantly enhancing the ability to distinguish between targets and backgrounds, and thus performing excellently in industrial scenarios with high noise.

[0098] To adapt to the limited computing resources in industrial sites, the large selection kernel crack detection model pays special attention to lightweight design. In the selection of convolutional kernels, the model widely uses 1×1 convolutional kernels instead of traditional large convolutional kernels, effectively reducing the computational complexity. In addition, by reducing redundant layers and optimizing convolutional operations, the storage requirements of the model are significantly reduced, enabling it to run efficiently on mobile or embedded devices. Even in resource-constrained environments, the model can still maintain a high level of detection performance.

[0099] The large selection kernel crack detection model of the present invention shows excellent performance in the defect detection task of sintered surfaces. For cracks of different severities (such as general cracks and extremely severe cracks), the detection accuracy and overall average precision of the model are significantly better than those of related detections in the prior art. In addition, the model is also significantly superior to traditional methods in terms of detection speed and robustness, and can adapt to various challenging scenarios such as light changes, complex textures, and industrial noise.

[0100] Step 2.1 Large selection kernel crack detection backbone network

[0101] Considering the complex texture of sintered surfaces and the characteristics of resource constraints in actual industrial scenarios, the backbone network design of the large selection kernel crack detection model focuses on reducing computational complexity and storage complexity while ensuring high detection accuracy. The backbone network is improved based on YOLOv7 and introduces multiple optimization modules to achieve the goals of lightweight and efficient feature extraction. The specific improvements include the following points:

[0102] 1. Modular feature extraction design

[0103] The backbone network extracts features from shallow to deep by combining CBS modules (convolution, normalization, Silu activation), CBS modules (convolution, normalization, Sigmoid activation), and ELAN modules (efficient hierarchical feature fusion). Among them, the CBS module balances the fineness of feature capture and computational efficiency through the combination of convolutional kernels of different sizes (such as 1×1 and 3×3); the ELAN module enhances the network's expression ability for multi-scale targets in complex scenarios through a multi-path feature fusion mechanism, while maintaining the efficient utilization of feature maps.

[0104] Among them, as Figure 4 shown, the CBS module extracts image features through convolution operations, and with the introduction of batch normalization and the Silu activation function, it improves the training stability and the non-linear expression ability of the model. As a result, when the network processes complex sintered surface images, it can effectively capture details and global features.

[0105] As Figure 5 shown, the CBM module adapts to the need for probability output in specific scenarios by changing the activation function from Silu to Sigmoid. For example, in the detection process, it is necessary to predict the confidence level of the target; Figure 4 、 Figure 5 in which k is the convolution kernel and s is the stride. Compared with the CBS module, the application of the Sigmoid function helps to enhance the performance of the model in certain specific tasks, especially in category probability prediction.

[0106] As Figure 6 shown, the ELAN module enhances the model's ability to express features at different scales in complex sintered surface images through feature fusion technology. This feature fusion method enables the large selection kernel crack detection to simultaneously focus on global information and local details, thus performing excellently in the detection tasks of complex backgrounds and diverse defects. The ELAN module effectively improves the robustness and accuracy of the model when facing multi-scales and multiple targets, where Figure 6 o and i represent different input sizes.

[0107] In this step, using the CBS module, CBM module, ELAN module, etc. as the main modules of the backbone network can improve the feature extraction efficiency, reduce the number of parameters and computational complexity while maintaining a high feature extraction accuracy, and strengthen the fusion of shallow features and deep features.

[0108] 2. Lightweight Structure Optimization

[0109] In terms of reducing the model complexity, the backbone network largely uses 1×1 convolutions to replace traditional 3×3 convolutions, reducing the number of parameters and the amount of computation. In addition, by modular design, redundant layers are reduced, further optimizing the storage and inference efficiency of the model, enabling it to be deployed on industrial field devices or mobile platforms with limited resources.

[0110] If the resolution of the input sintered image is 640×640, the processing flow of the backbone network is roughly as follows:

[0111] (1) The input image first undergoes feature extraction by multiple CBS and ELAN modules, gradually generating multi-scale feature maps (C3, C4, C5).

[0112] (2) After passing through the LSKNet module, the feature map extracts high-dimensional feature information with a dynamic receptive field.

[0113] (3) Finally, multi-scale pooling is performed through the SPPCSPC module to effectively fuse global and local information, providing accurate feature representations for the detection head.

[0114] Through the above design, the backbone network of the large selection kernel crack detection model performs excellently in sintered surface target detection, not only ensuring the detection accuracy, but also having good real-time and lightweight characteristics, providing technical support for the efficient deployment of the intelligent smelting system.

[0115] Step 2.2 Large Selection Kernel Network Module (LSKNet)

[0116] The LSKNet module is one of the core innovations in the large selection kernel crack detection model, used to dynamically adjust the receptive field to adapt to the multi-scale feature requirements in sintered surface defect detection tasks. This module effectively expands the receptive field range through a large convolution kernel selection mechanism while maintaining the controllability of the model's computational complexity.

[0117] The LSKNet includes the neural network module design of LSK Selection and FFN (feed-forward network), mainly used to enhance the network's feature extraction and modeling capabilities. As Figure 7 shown, the module starts with two Norm normalization operations, and then is processed through LSK Selection and FFN. LSK Selection focuses on dynamically selecting features of different scales, calling the upper right branch structure, that is, from the fully connected layer (FC) to the GELU activation, then to the LSK module, and finally to the FC, to achieve the feature extraction of the large kernel receptive field. The FFN part combines the lower right branch, that is, from the FC to the depthwise separable convolution (DW-Conv), then to the GELU activation, and finally to the FC, to achieve efficient local feature extraction and non-linear transformation.

[0118] The key design of the LSKNet module lies in the combination of a large convolution kernel and a selection mechanism. In the module, a large receptive field is gradually constructed through multiple layers of depth convolution, and the dilated convolution strategy is adopted to expand the receptive field range without increasing the number of parameters. This design can establish a connection between global and local features, especially suitable for processing sintered surface images containing complex backgrounds and diverse targets. In addition, the spatial selection mechanism within the module dynamically selects convolution kernels of different sizes, thereby preferentially retaining relevant features and suppressing redundant information in specific scenarios, effectively improving the detection accuracy and robustness.

[0119] To further optimize the computational efficiency, the LSKNet module introduces depthwise separable convolutions, which decompose the standard convolution into a depthwise convolution and a pointwise convolution (i.e., 1×1 convolution). The depthwise convolution independently performs spatial convolutions on each input channel, thus preserving spatial features, while the pointwise convolution integrates the output of the depthwise convolution through linear combinations between channels. This method significantly reduces the number of parameters and computational complexity, enabling the LSKNet module to be not only applicable to high-precision detection scenarios but also to operate efficiently in resource-constrained industrial devices.

[0120] Through the combination of the large convolution kernel selection mechanism and depthwise separable convolutions, the LSKNet module achieves a good balance between expanding the receptive field and improving the detection efficiency. Its dynamic selection ability ensures that the model performs excellently in dealing with the multi-scale defect detection task of the sintered surface, providing strong feature expression support for the large selection kernel crack detection model.

[0121] Step 2.3 Pyramid Pooling Module Design

[0122] In the object detection task of the sintered surface, the defects in the sintered surface image often exhibit a diverse scale distribution, including both cracks with a large coverage area and fine surface defects. Traditional object detection models often struggle to simultaneously consider the global information of large-scale objects and the detailed features of small-scale objects, resulting in insufficient detection accuracy. To address this issue, the large selection kernel crack detection model designs and introduces a pyramid pooling module (SPPCSPC) to enhance the multi-scale feature extraction ability.

[0123] Refer to Figure 8 As shown, the SPPCSPC module realizes the feature modeling of different-scale objects through multi-resolution pooling operations on the feature map. Specifically, the module uses pooling kernels of sizes 1×1, 5×5, 9×9, and 13×13 to perform max-pooling operations on the input feature map, generating feature maps with different receptive field ranges. Subsequently, the pooled multi-scale feature maps are integrated into a single feature representation through a fusion operation, thus achieving a balance between global and local information. This multi-scale feature representation ensures that the model can simultaneously handle large-scale cracks and fine surface defects.

[0124] Compared with traditional fixed-scale feature extraction methods, the SPPCSPC module not only dynamically adapts to the sizes of different objects in feature map generation but also simplifies the overall architecture of the model by reducing the complex process of multi-stage candidate region generation, improving the speed and efficiency of object detection. In addition, the module retains key contextual information through the pyramid-style pooling design, avoiding information loss of small-scale objects during the feature fusion process, thus significantly improving the detection sensitivity and accuracy.

[0125] By introducing the pyramid pooling module, the large selection kernel crack detection model can effectively address the challenges brought about by the changing target scales in sintered surface images, achieve efficient adaptation to complex industrial scenarios, and further improve the overall performance of defect detection.

[0126] Specifically, the spatial pyramid pooling module includes:

[0127] 1) One 1×1 convolutional layer and three 3×3 dilated convolutional layers for lightweight multi-scale feature extraction;

[0128] 2) One global average pooling layer to obtain image-level features, which are then fed into a 1×1 convolutional layer, and bilinear interpolation is used to obtain an image of the same size;

[0129] 3) The five multi-scale features obtained are merged together through the same channel dimension, and then fed into a 1×1 convolution for post-fusion to obtain feature information with a new number of channels.

[0130] The designed pyramid pooling module can generate feature information of different scales and achieve feature fusion. After obtaining the deep features, they are input into the decoder. This module includes a top-down module, a bottom-up module, and a fusion and expansion module to perform feature learning and fusion, realizing the simultaneous detection of large ranges and small targets, and well solving the problem of feature extraction for sintered images of different specifications. At the same time, this module can avoid losing the context information representing the relationship between different sub-regions.

[0131] In step 2, an optimized YOLO detection head is adopted, combined with high-resolution and low-resolution feature maps, to perform multi-level feature map prediction, realizing the localization and classification of targets of different scales; the regression method is used to accurately predict the bounding box coordinates and class confidence of the defect area.

[0132] Step 3: Training and prediction of the target model;

[0133] In the training and prediction stages of the large selection kernel crack detection model, the labeled sintered surface defect image dataset is divided into a training set, a validation set, and a test set. When the data volume is sufficient, it can be divided according to the ratio of 7:2:1 to ensure that the training set can provide sufficient data support for model learning, while the validation set and the test set can be used to evaluate the generalization performance and actual detection ability of the model.

[0134] During the training process, the model uses the training set for multiple rounds of iterative optimization, gradually improving the accuracy of object detection by adjusting the weight parameters. To enhance the robustness and adaptability of the model, data augmentation strategies are also adopted during the training phase, such as random cropping, color transformation, and image rotation, etc., to simulate diverse inputs in real industrial scenarios. The validation set is used to evaluate the performance of the model in real-time during the training process, so as to adjust the hyperparameters according to the validation results and prevent the occurrence of overfitting problems.

[0135] In the testing phase, the finally trained large selection kernel crack detection model is applied to the test set to evaluate its actual detection effect.

[0136] Step 4: Real-time defect detection;

[0137] Considering the actual requirements of the industrial site, when the accuracy of the model in the target defect detection task reaches the expectation (such as higher than 95%), it can be applied to the real-time monitoring and defect location of the sintering surface quality.

[0138] Perform real-time defect detection on the mineral sintering surface image to be detected, output the grade of each crack, and calculate the overall severity score of the surface based on the grade of each crack. The specific calculation process refers to the calculation steps of the sinter ore surface defect severity score S in Step 1 above, which will not be elaborated here.

[0139] Through efficient learning of the dataset, the large selection kernel crack detection model shows good adaptability to the scenario with insufficient labeled data, and at the same time takes into account high detection performance and lightweight design, capable of meeting the requirements of the intelligent smelting system for an efficient and reliable detection solution.

[0140] The lightweight surface crack detection method for mineral sintering surfaces proposed by the present invention has the following remarkable beneficial effects:

[0141] 1. Improve detection accuracy: By combining the texture features and semantic information of the sintering surface image, this method designs a deep neural network model to achieve precise detection of complex surface defects. Compared with traditional machine vision methods and general object detection models, this method has stronger adaptability to low-contrast, high-noise, and complex texture environments, significantly improving the detection accuracy.

[0142] 2. Achieve model lightweight: Aiming at the characteristics of limited resources in the industrial site, this method optimizes the network structure, significantly reducing the computational complexity and storage requirements. The lightweight design is not only applicable to small and medium-sized devices and mobile deployments, but also can save resource costs while ensuring high detection performance.

[0143] 3. Enhanced generalization: This method is specifically designed for the defect characteristics of the mineral sintering surface. By introducing a multi-scale feature extraction mechanism and dilated convolution technology, it effectively improves the adaptability to environmental light changes, moving probe interference, and other complex working conditions, thus showing high stability and generalization in various actual scenarios.

[0144] 4. Meeting real-time requirements: This method focuses on optimizing the detection speed in the model structure design. Through an efficient feature extraction and processing process, it realizes the real-time response ability for defect detection. This feature can meet the requirements of the industrial field for rapid feedback and provides a guarantee for the improvement of production efficiency.

[0145] 5. Promoting the development of intelligent smelting: This invention provides a reliable defect detection technology for the intelligent smelting system, laying a foundation for the automation, intelligentization, and green development of the steel industry. By providing real-time feedback of defect information, it can assist engineers in quickly adjusting production parameters, thereby reducing resource consumption, environmental pollution, and improving the overall efficiency of the sintering process.

[0146] In summary, this invention effectively solves the deficiencies of existing methods in the detection of mineral sintering surface defects, combines the characteristics of high precision, lightweight, and high generalization, and has broad application prospects.

[0147] Embodiment 2:

[0148] Based on the same inventive concept, an embodiment of the present invention also provides a lightweight surface crack detection system for the mineral sintering surface. As shown in Figure 9 it includes:

[0149] A dataset construction module for segmenting and classifying the cracks in the collected mineral sintering surface images to construct a sintered ore surface defect dataset; dividing the sintered ore surface defect dataset into a training set, a validation set, and a test set;

[0150] A detection module construction module for constructing a large selection kernel crack detection model based on the improved YOLOv7. The model includes:

[0151] (a) A backbone network composed of multiple CBS modules, CBM modules, and ELAN modules for multi-scale feature extraction;

[0152] (b) An LSKNet module that realizes the fusion of global and local features by dynamically adjusting the receptive field range and combining depthwise separable convolution and a spatial selection mechanism;

[0153] (c) An SPPCSPC module that generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintering surface;

[0154] A training and testing module for training the model using a training set labeled with crack classifications, optimizing the model's robustness through data augmentation strategies, adjusting hyperparameters based on a validation set, and evaluating the detection effect based on a test set;

[0155] A detection module that uses the trained model to perform real-time defect detection on the surface image of the mineral sinter to be detected, outputs the level of each crack, and calculates the overall severity score of the surface based on the level of each crack.

[0156] Through the collaborative work of the above modules, and through innovative model design and optimization, the system realizes precise and rapid detection of cracks on the surface of mineral sinter, has broad application prospects, and can effectively promote the intelligent development of the steel industry.

[0157] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0158] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightweight surface crack detection method for a mineral sintering surface, characterized in that, It includes the following steps: S10. Segment the collected mineral sintering surface images and classify the cracks to construct a sintered ore surface defect dataset; Divide the sintered ore surface defect dataset into a training set, a validation set and a test set; S20. Construct a large selection kernel crack detection model based on the improved YOLOv7. The model includes: (a) A backbone network, which is composed of multiple CBS modules, CBM modules and ELAN modules for multi-scale feature extraction; (b) An LSKNet module, which realizes the fusion of global and local features by dynamically adjusting the receptive field range and combining depthwise separable convolution and spatial selection mechanism; (c) An SPPCSPC module, which generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintering surface; S30. Train the model with the training set labeled with crack classification, optimize the model robustness through data augmentation strategies, adjust the hyperparameters based on the validation set, and evaluate the detection effect based on the test set; S40. Use the trained model to perform real-time defect detection on the mineral sintering surface images to be detected, output the level of each crack, and calculate the overall severity score of the surface based on the level of each crack.

2. The lightweight surface crack detection method for a mineral sintering surface according to claim 1, wherein, The crack classification in S10 includes: Divided into 1-4 categories according to the ratio of crack length to burden surface width.

3. A lightweight surface crack detection method for a mineral sintered surface according to claim 1, characterized in that, The CBS module includes a convolutional layer, a batch normalization layer and a Silu activation function connected in sequence, which is used to balance the fineness of feature crack capture and the calculation efficiency; The CBM module includes a convolutional layer, a batch normalization layer and a Sigmoid activation function connected in sequence, which is used to predict the confidence of the target crack; The ELAN module enhances the expression ability of different scale features through a multi-path feature fusion mechanism.

4. A lightweight surface crack detection method for a mineral sintering surface according to claim 1, characterized in that, The backbone network also includes: Lightweight design, using 1×1 convolution kernels to replace traditional large convolution kernels and reducing redundant layers to reduce computational complexity.

5. A lightweight surface crack detection method for a mineral sintering surface according to claim 1, characterized in that, The LSKNet module includes: a) An LSK Selection branch, which adopts a large convolution kernel selection mechanism to realize the feature extraction of the large kernel receptive field through a fully connected layer and a GELU activation function; b) An FFN branch, which combines depthwise separable convolution and a GELU activation function to realize efficient local feature extraction.

6. The lightweight surface crack detection method for a mineral sintering surface according to claim 1, characterized in that, The SPPCSPC module includes: 1) 1 1×1 convolutional layer and 3 3×3 dilated convolutional layers for lightweight multi-scale feature extraction; 2) 1 global average pooling layer to obtain image-level features, and then send them into a 1×1 convolutional layer, and use bilinear interpolation to obtain an image of the same size; 3) Merge the obtained 5 multi-scale features together through the same channel dimension, and then send them into a 1×1 convolution for post-fusion to obtain feature information with a new channel number.

7. A lightweight surface crack detection method for a mineral sintering surface according to claim 1, characterized in that The data augmentation strategies include random cropping, color transformation and image rotation to simulate the illumination changes and noise interference in the industrial scenario.

8. A lightweight surface crack detection method for a mineral sintering surface according to claim 1, characterized in that, In step S40, calculating the overall severity score of the surface based on the level of each crack includes: The formula for the score of a single crack is as follows: Among them, the crack level i is divided into levels 1-4 according to the ratio of the crack length to the burden surface width, corresponding to different reference scores and coefficients respectively; Where S represents the severity score of the overall surface defects; S k represents the crack score of the k-th crack; K represents the total number of cracks in the image.

9. A lightweight surface crack detection system for a mineral sintering surface, characterized in that, including: a dataset construction module, configured to segment and classify cracks in the collected mineral sintering surface images, and construct a sinter ore surface defect dataset; divide the sinter ore surface defect dataset into a training set, a validation set and a test set; a detection module construction module, configured to construct a large selection kernel crack detection model based on the improved YOLOv7, and the model includes: (a) a backbone network, composed of a plurality of CBS modules, CBM modules and ELAN modules, for multi-scale feature extraction; (b) an LSKNet module, which realizes the fusion of global and local features by dynamically adjusting the receptive field range and combining depthwise separable convolution and spatial selection mechanism; (c) an SPPCSPC module, which generates feature maps with different resolutions through multi-scale pooling operations, enabling the model to adapt to defect targets of different sizes on the sintering surface; a training and testing module, configured to train the model with the training set with labeled crack classification, optimize the model robustness through data augmentation strategies, adjust hyperparameters based on the validation set, and evaluate the detection effect based on the test set; a detection module, which uses the trained model to perform real-time defect detection on the mineral sintering surface image to be detected, outputs the crack level of each line, and calculates the overall severity score of the surface based on the crack level of each line.

Images