An intelligent detection method and system for steel surface defect identification

Abstract

The present invention discloses an intelligent detection method for steel surface defect identification, comprising the following steps: S1, image acquisition: real-time image acquisition, setting trigger conditions according to needs, taking photos at intervals or triggering with external signals, and capturing images at specified times; S2, image preprocessing: preprocessing the acquired images, including image scaling and color space conversion steps; S3, defect detection: inputting the preprocessed images into a target detection model for inference, and outputting detection results including the category and bounding box coordinates of each target; S4, result display and positioning: drawing bounding boxes and category labels on the original image to visualize the detection results. The present invention also discloses an intelligent detection system for steel surface defect identification. The present invention aims to solve the technical problems in accuracy and efficiency in the process of steel surface defect detection, and can be widely used in the field of deep learning technology.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an intelligent detection method and system for identifying surface defects of steel. Background Art

[0002] Steel is crucial in modern industry. The equipment, processing parameters, and raw material quality during its production process all affect the performance of the final product. Surface defects such as cracks and scratches can reduce its safety. Existing technologies primarily include traditional manual inspection methods and eddy current testing technology.

[0003] Traditional manual inspection methods rely on the experience and skills of inspectors, usually performed through visual inspection or the use of handheld inspection equipment. This method is slow, inefficient, and prone to missed or false detections.

[0004] Eddy Current Testing (ECT) is a non-destructive testing technology based on the principle of electromagnetic induction. When an alternating current is passed through the coil inside the eddy current probe, an alternating magnetic field is generated. When close to a conductive material, this magnetic field will generate eddy currents inside the material. The intensity and distribution of eddy currents are affected by the electrical conductivity, magnetic permeability and geometric shape of the material. If there are defects in the material (such as cracks and holes), the characteristics of the eddy current will be changed, thereby affecting the electrical signal sensed by the probe. By processing and analyzing these signals, defects in the material can be identified. ECT has high sensitivity and can detect tiny defects on the surface of steel. It can make up for the problems of traditional manual detection methods to a certain extent and improve detection efficiency and quality.

[0005] Eddy current testing is a non-destructive testing method widely used in metal material inspection, primarily for detecting surface and near-surface defects such as cracks, inclusions, and delamination. Defects in the material, such as cracks or holes, disrupt the normal distribution of eddy currents, altering their paths, intensity, and distribution. Eddy current testing boasts high sensitivity and rapid detection speed, making it widely used in industrial production. However, it has significant drawbacks: First, the choice of testing frequency affects detection effectiveness. Generally speaking, high-frequency eddy currents can detect surface defects, while low-frequency eddy currents can detect deeper defects. The selection of testing frequency often relies on human experience, which can easily lead to errors in test results. Second, eddy current testing is susceptible to material thickness, shape, and size limitations. Excessively thick or thin materials can affect detection accuracy and stability, and very small or very large defects may be missed or falsely detected. Summary of the Invention

[0006] The present invention aims to overcome the shortcomings of the aforementioned background technology by providing an intelligent detection method and system for steel surface defect identification, which aims to solve the technical problems of accuracy and efficiency in the steel surface defect detection process. By applying intelligent visual recognition technology, steel surface defects can be detected more accurately and quickly, improving detection efficiency and accuracy.

[0007] The present invention provides an intelligent detection method for identifying surface defects in steel, comprising the following steps: S1, image acquisition: real-time image acquisition, setting trigger conditions according to requirements, taking photos at a specified time or triggering by an external signal, and capturing images at a specified time; S2, image preprocessing: preprocessing the acquired image, including image scaling and color space conversion steps; S3, defect detection: inputting the preprocessed image into a target detection model for inference, and outputting detection results including the category and bounding box coordinates of each target; S4, result display and positioning: drawing bounding boxes and category labels on the original image to visualize the detection results.

[0008] In the above technical solution, the specific process of step S3 is as follows: S31, target detection model construction: construct a model consisting of a backbone network, a neck network and a detection head; the backbone network introduces a deep separable convolutional network to extract features; the neck network adopts the C3 module and the bidirectional feature pyramid network to fuse features, wherein the C3 module filters background interference and the bidirectional feature pyramid network enhances the expression ability of features at each scale; the detection head constructs the final target classification and bounding box regression; S32, target detection model training: input the preprocessed image data into the target detection model for training, and the target detection model training adopts the large-scale open source NEU surface defect detection dataset; S33, target detection model reasoning: input the real-time collected image into the trained model, and the model calculates the prediction results of each grid unit, including the target category, bounding box position and confidence. After non-maximum suppression, duplicate detection results are removed, the bounding box with the highest confidence is retained, and the category and bounding box coordinates of each detected target are output.

[0009] In the above technical solution, in step S31, the specific process of introducing the deep separable convolutional network to extract features in the backbone network is as follows: input feature map convolution operation: use N convolution kernels with a size of S×S and a depth of M to perform a convolution operation on the input feature map to obtain N H×W feature maps. The parameter amount of the convolution operation is: P1=S×S×M×N; deep separable convolutional network operation: use M convolution kernels with a size of S×S and a depth of 1 to perform a deep convolution operation on the input feature map. Convolution is performed point by point to obtain M H×W feature maps, and the parameter quantity is: P2 = S×S×1×M; then N convolution kernels of size 1×1 and depth M are used to perform point-by-point convolution on the output of the convolution operation of the input feature map to generate N H×W feature maps, and the parameter quantity is: P3 = 1×1×M×N. The total parameter quantity of the depthwise separable convolutional network operation is: P4 = S×S×1×M+1×1×M×N. The ratio of the parameter quantity of the depthwise separable convolutional network operation to that of the ordinary convolution is:

[0010]

[0011] In the above technical solution, in step S31, the specific working process of the C3 module in the neck network is as follows: the C3 module adopts a SiLU structure, including a main branch and an auxiliary branch, wherein the main branch includes a convolution model and a Bottleneck block model, and the auxiliary branch includes a convolution model. The output obtained by the auxiliary branch is connected with the result calculated by the main branch, and then the feature information is further integrated through a convolution model again. The three convolution models all perform a 1×1 convolution operation.

[0012] In the above technical solution, in step S31, the specific working process of the Bottleneck block model of the C3 module in the neck network is as follows: the Bottleneck block model includes two branches, wherein the first branch first halves the number of channels of the feature map through 1×1 convolution, and then extracts features through 3×3 convolution and doubles the number of channels; the second branch performs residual connection through shortcut, adds the original input to the feature map after convolution operation, and realizes feature fusion.

[0013] In the above technical solution, in step S31, the specific working process of the bidirectional feature pyramid network in the neck network is as follows: first, the input nodes with low contribution to feature fusion are removed; second, the bidirectional feature pyramid network locates and identifies targets at various scales by introducing cross-scale connections to connect features from various scales; third, a weighted feature fusion mechanism is introduced to fuse the features of each scale after connection, and the features at each level are assigned corresponding weights. These weights are automatically learned during the training process to ensure the best integrated features.

[0014] In the above technical solution, in step S32, when the target detection model trains the open source NEU surface defect detection dataset, the target detection model optimizes parameters by minimizing the loss function, and the loss function includes classification loss, bounding box regression loss and confidence loss.

[0015] The above technical solution further includes step S5, defect type warning: storing the detection results and sending warnings for defects that reach a threshold.

[0016] In the above technical solution, step S6 and process control are also included: the controller determines whether the image sensor reaches the movement threshold. If not, it moves a specified distance and continues detection; if it reaches, it requests a reset instruction; if reset is selected, the new steel sample detection continues; if not reset is selected, the detection process ends.

[0017] The present invention also provides an intelligent detection system for identifying surface defects of steel materials, which has a computer program capable of executing the intelligent detection method for identifying surface defects of steel materials.

[0018] The intelligent detection method and system for steel surface defect identification of the present invention have the following features:

[0019] Beneficial effects:

[0020] (1) The core part of the present invention adopts machine vision technology, which performs automated detection through preset algorithms and models, and can maintain consistent detection standards under different detection conditions, ensuring the repeatability and reliability of each detection and reducing interference from human factors.

[0021] (2) The present invention can capture rich information in the image, including color, shape, texture, etc., which can better identify and classify various types of defects and improve the accuracy of detection. In contrast, eddy current detection mainly relies on electromagnetic signals and is difficult to provide comprehensive defect characteristics.

[0022] (3) The data preprocessing step and defect detection step in the present invention can be flexibly adapted to the detection of products of different materials, shapes and sizes by adjusting the algorithm and parameters; however, eddy current detection is limited by the thickness and shape of the material and is difficult to adapt to diverse detection needs.

[0023] (4) The present invention can quickly process a large number of images and provide timely feedback on test results. Real-time detection makes it suitable for online testing on production lines. Eddy current testing usually requires a long preparation and adjustment time, especially in frequency selection and equipment calibration.

[0024] (5) Wide range of detection types: The present invention can detect six types of steel defects, including foreign matter inclusions, scratches, cracks, plaques, rolling oxidation and surface pitting, while eddy current detection is mainly suitable for specific defect types (such as cracks in conductive materials). BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the functions of the intelligent detection method and system for identifying steel surface defects according to the present invention;

[0026] Figure 2 Schematic diagram of the overall process of the intelligent detection method for steel surface defect identification according to the present invention;

[0027] Figure 3 Schematic diagram of the process of step S4 of the intelligent detection method for identifying surface defects of steel materials according to the present invention;

[0028] Figure 4 Schematic diagram of the workflow of each algorithm in step S41 of the intelligent detection method for steel surface defect identification of the present invention;

[0029] Figure 5 Schematic diagram of the topological structure of the C3 module in step S412 of the intelligent detection method for steel surface defect identification according to the present invention;

[0030] Figure 6 Schematic diagram of the topological structure of the Bottleneck block in the C3 module of step S412 of the intelligent detection method for steel surface defect identification according to the present invention;

[0031] Figure 7 Schematic diagram of the workflow of the bidirectional feature pyramid network in step S413 of the intelligent detection method for steel surface defect identification of the present invention;

[0032] Figure 8 Schematic diagram of six main defects on the surface of steel materials included in the large-scale open source NEU surface defect detection dataset in step S42 of the intelligent detection method for steel surface defect identification of the present invention;

[0033] Figure 9 Schematic diagram of the architecture of the intelligent detection system for steel surface defect identification according to the present invention;

[0034] Figure 10 Schematic diagram of the results of qualitative comparison between the intelligent detection method and system for steel surface defect identification and other deep learning detection algorithms. DETAILED DESCRIPTION

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and examples, but the examples should not be construed as limiting the present invention.

[0036] Intelligent machine vision-based inspection technology (the method related to this invention) utilizes image processing and computer vision algorithms to automatically detect defects on product surfaces. The basic principle is to capture images of the object being inspected using a high-resolution camera and then analyze them using image processing software. The intelligent algorithm unit (i.e., the defect detection unit) identifies the location and type of defects through feature extraction, image segmentation, and pattern recognition.

[0037] This invention requires detection and analysis equipment, primarily consisting of a data acquisition unit and a system host. The data acquisition unit, primarily a camera (i.e., image sensor), captures images of the steel surface; the system host integrates data processing and intelligent algorithms. The data processing function preprocesses the captured images and detects surface features, while the intelligent algorithm identifies and analyzes defects. This technical solution improves detection accuracy and efficiency while reducing manpower and costs.

[0038] 1. Technical functions (see Figure 1 )

[0039] Image sensor function: Used to collect image information of the steel surface. The sensor module includes two types: RGB camera and depth camera.

[0040] Data processing function: Receive images collected by the sensor and perform cropping and filtering on the signals to extract useful information and reduce noise interference.

[0041] Intelligent algorithm function: An innovative target detection algorithm is used to analyze and process the collected data to identify and classify defects on the steel surface.

[0042] Display and alarm function: The test results are displayed on the human-machine interface, and the corresponding alarm mechanism is triggered according to the test results to remind the operator to carry out subsequent processing.

[0043] 2. Technical solution

[0044] See also Figure 2 The workflow of the intelligent detection method for steel surface defect identification of the present invention can be divided into the following steps:

[0045] S1. MCU Initialization: Hardware Preparation: Ensure the camera module is correctly installed and properly connected to the computing platform. Configure camera parameters such as resolution, frame rate, and exposure time to ensure high-quality capture.

[0046] S2, Image Acquisition: Start the motor and begin real-time image acquisition. Set the trigger conditions as needed, and choose to capture images at a specified time using timed photography or external signal triggering.

[0047] S3. Image preprocessing: Preprocess the collected images to improve the effectiveness of subsequent detection. This includes image scaling and color space conversion (from BGR to RGB).

[0048] S4. Defect Detection: The preprocessed image is fed into the object detection model for inference. The model outputs the detection results, including the category and bounding box coordinates of each object.

[0049] The defect detection step, i.e. the intelligent algorithm function, is the core of this invention, which is equipped with a new and improved YOLOv9 network. The construction principle and process of this step are introduced below. Figure 3 :

[0050] S41. Object Detection Model Construction: This model consists of three main components: the backbone network (Backbone), the neck network (Neck), and the detection head (Head). The backbone network is responsible for feature extraction and uses a deep separable convolutional network based on the YOLOv9 architecture. The neck network is used for feature fusion, incorporating a C3 module and a bidirectional feature pyramid network to filter background interference and enhance the representation of features at different scales, respectively. The detection head is responsible for final object classification and bounding box regression.

[0051] The algorithm embedded in step S41 is as follows. Figure 4 :

[0052] S411. The deep separable convolutional (DWConv) network is introduced into the backbone network structure to reduce redundant feature information, reduce the number of parameters of the target detection model, and speed up the inference speed of the target detection model.

[0053] In one or more embodiments,

[0054] The input feature map is convolved using N convolution kernels of size S×S and depth M to obtain N feature maps of size H×W. The number of parameters in this process is: P1 = S×S×M×N.

[0055] The process of depth-wise separable convolution (DWConv): First, the input feature map is convolved with M convolution kernels of size S×S and depth 1 to obtain M H×W feature maps. The parameter amount is: P2 = S×S×1×M; then, N convolution kernels of size 1×1 and depth M are used to perform point-by-point convolution on the result of the previous step to generate N H×W feature maps. The parameter amount is: P3 = 1×1×M×N. The total number of DWConv parameters is: P4 = S×S×1×M+1×1×M×N. The ratio of the number of parameters of DWConv and ordinary convolution is:

[0056]

[0057] In formula ①, the ratio of P4 to P1 is much less than 1, indicating that the use of DWConv can significantly reduce the complexity of the model, reduce the computational cost, and improve the detection speed of the model.

[0058] S412. The C3 module is introduced into the neck network structure, which helps to filter background interference, allowing the target detection model to focus more on the target area of ​​interest, retaining key details in the image and producing clearer and more accurate output.

[0059] As an innovative component based on a deep learning architecture, the C3 module consists of three convolutional modules and n Bottleneck modules. The specific composition is as follows:

[0060] The main branch contains a Conv block and a Bottleneck block, while the auxiliary branch contains only one Conv block. The output is connected with the result of the main branch, and then further integrated with the feature information through another Conv block. The Conv blocks here all perform 1×1 convolution operations. In addition, the C3 module uses SiLU (Sigmoid-like Unit) instead of the traditional LeakyReLU. The topology of the C3 module is as follows: Figure 5 shown.

[0061] The following is a detailed explanation of the Bottleneck block in the C3 module. The Bottleneck block also contains two branches. The first branch first reduces the number of channels of the feature map by half through 1×1 convolution, and then extracts features through 3×3 convolution and doubles the number of channels. The second branch performs residual connection through shortcut, adds the original input and the feature map after convolution operation, and realizes feature fusion. The topological structure of the Bottleneck block is as follows: Figure 6 shown.

[0062] S413. In the YOLOv9-based neck network structure, a bidirectional feature pyramid network (BiFPN) is introduced. By enhancing the target detection model's recognition of the geometric shape of the target in the image and its understanding of the scene environment, the target detection model can more accurately identify the edge and shape of the target, reducing missed detections and false alarms. For details, see Figure 7 :

[0063] The difference between BiFPN and the existing spatial pyramid structure is that: first, some input nodes with low contribution to feature fusion are removed, such as Figure 7The P3 and P7 nodes are shown. This improvement not only improves the efficiency of feature fusion but also reduces network parameters and computational load. Secondly, by introducing cross-scale connections, BiFPN is able to fuse features from different scales, thereby accurately localizing and identifying objects at different scales. During the feature fusion process, a weighted feature fusion mechanism is introduced, in which features at different levels are assigned different weights. These weights are automatically learned during training to ensure optimal feature integration.

[0064] S42. Object detection model training: Input the prepared dataset into the object detection model for training. The large-scale open-source NEU surface defect database (NEU-DET) is used for model training.

[0065] Among them, NEU-DET includes six main defects that appear on the surface of steel. They are inclusions (In), scratches (Sc), cracks (Cr), patches (Pa), rolled-in scale (Rs) and pitted surface (Ps). Different defect samples such as Figure 8 shown.

[0066] During the NEU-DET training process, the target detection model optimizes parameters by minimizing the loss function, which includes classification loss, bounding box regression loss, and confidence loss.

[0067] S43, Object Detection Model Inference: New images captured by the camera module are fed into the trained object detection model. The model calculates a prediction for each grid cell, including the target's category, bounding box location, and confidence level. Non-maximum suppression (NMS) is used to remove duplicate detection results and retain the bounding box with the highest confidence level. Finally, the model outputs the category and bounding box coordinates for each detected target.

[0068] S5. Result display and positioning: Draw bounding boxes and category labels on the original image to intuitively visualize the detection results.

[0069] S6. Defect type warning: This step will store the detection results and send a warning when the defect threshold (length × width) is reached. The system will track further analysis and processing results.

[0070] S7, Process Control: The MCU determines whether the camera module has reached the movement threshold. If not, it moves the specified distance and continues testing. If it has, it requests a reset instruction. If reset is selected, the new steel sample is tested. If not, the process ends.

[0071] See also Figure 9 The intelligent detection system for steel surface defect identification of the present invention includes the following parts:

[0072] Controller initialization unit: Hardware preparation ensures that the camera module is correctly installed and properly connected to the computing platform. Configure camera parameters such as resolution, frame rate, exposure time, etc. to ensure shooting quality;

[0073] Image acquisition unit: real-time image acquisition, set trigger conditions according to needs, take photos at regular intervals or trigger with external signals to capture images at the specified time;

[0074] Image preprocessing unit: preprocesses the collected images, including image scaling and color space conversion steps;

[0075] Defect detection unit: Inputs the preprocessed image into the object detection model for inference, and outputs the detection results including the category and bounding box coordinates of each object;

[0076] Result display and positioning unit: draw bounding boxes and category labels on the original image to visualize the detection results;

[0077] Defect type warning unit: stores detection results and sends warnings for defects that reach the threshold;

[0078] Process control unit: The controller determines whether the image sensor has reached the movement threshold. If not, it moves the specified distance and continues the detection; if it has reached, it requests a reset instruction; if reset is selected, the new steel sample detection continues; if not reset is selected, the detection process ends.

[0079] 3. Comparison of the technical solution of the present invention with other defect detection algorithms

[0080] With the development of the field of computer vision, there are many defect detection technologies based on deep learning. Different neural networks have different advantages and are suitable for images in different fields and with different characteristics. During the research and development of the present invention, comparative experiments were conducted on a variety of existing solutions. The existing methods can also achieve the task of steel surface defect detection to a certain extent, but they still have different shortcomings. For example, some models have too many parameters and too slow inference time, which are not suitable for deployment in embedded devices; some models have too high a missed detection rate or false alarm rate for a certain type of defect; some models have too poor feature extraction capabilities for depth maps, etc. The following table is the experimental results obtained after quantitative comparison of various defect detection algorithms. At the same time, the results of qualitative comparison of the present invention with other deep learning detection algorithms can be found in Figure 10 .

[0081] Quantitative comparison results of different defect detection algorithms

[0082]

[0083] 4. Terminology

[0084] MCU——Microcontroller Unit, micro control unit.

[0085] YOLO — You Only Look Once, a deep learning object detection framework.

[0086] BGR, RGB - the three channels of a color image, B stands for Blue, G stands for Green, and R stands for Red.

[0087] Precision: Accuracy P,

[0088] AP: Average Precision,

[0089] mAP: mean Average Precision,

[0090] Among them, T p Indicates the number of defect targets detected correctly; F p Indicates the number of defect targets that are falsely detected; F N The number of missed defect targets; n represents the number of defect categories; AP(i) is the average precision representing the i-th target class.

[0091] FPS: (Frames Per Second) indicates the number of image frames displayed per second. An index greater than 35FPS can meet real-time requirements.

[0092] GFLOPs: (Giga Floating-point Operations Per Second, 1 billion floating-point operations per second), often used as a GPU performance parameter to measure algorithm / model complexity.

[0093] Parameters: parameter quantity.

[0094] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0095] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An intelligent detection method for identifying surface defects of steel, characterized by: The steps include: S1. Image acquisition: Real-time image acquisition, set trigger conditions according to needs, take photos at regular intervals or trigger with external signals, and capture images at the specified time; S2, image preprocessing: preprocessing the collected images, including image scaling and color space conversion steps; S3, defect detection: The preprocessed image is input into the object detection model for inference, and the output includes the detection results of each object's category and bounding box coordinates. The specific process is as follows: S31. Target detection model construction: A model consisting of a backbone network, a neck network, and a detection head is constructed. The backbone network introduces a deep separable convolutional network based on the YOLOv9 structure to extract features. The neck network uses the C3 module and a bidirectional feature pyramid network to fuse features. The C3 module filters background interference, and the bidirectional feature pyramid network enhances the expressiveness of features at each scale. The detection head performs the final target classification and bounding box regression. The specific process of introducing a deep separable convolutional network to extract features in the backbone network is as follows: Convolution operation on the input feature map: Use N convolution kernels of size S×S and depth M to perform convolution operation on the input feature map to obtain N H×W feature maps. The parameter quantity of this convolution operation is: P1=S×S×M×N; Depthwise separable convolutional network operation: Through M convolution kernels of size S×S and depth 1, the input feature map is depthwise convolved to obtain M H×W feature maps. The parameter amount is: P2=S×S×1×M; then N convolution kernels of size 1×1 and depth M are used to perform point-by-point convolution on the output of the convolution operation of the input feature map to generate N H×W feature maps. The parameter amount is: P3=1×1×M×N, so the total parameter amount of the depthwise separable convolutional network operation is: P4=S×S×1×M+1×1×M×N, and the parameter amount ratio of the depthwise separable convolutional network operation to the ordinary convolution is: ； S32. Target detection model training: The preprocessed image data is input into the target detection model for training. The target detection model training uses the large-scale open source NEU surface defect detection dataset; S33, Object Detection Model Inference: The real-time captured image is input into the trained model. The model calculates the prediction results for each grid cell, including the target category, bounding box location, and confidence level. After non-maximum suppression, duplicate detection results are removed, the bounding box with the highest confidence level is retained, and the category and bounding box coordinates of each detected target are output. S4. Result display and positioning: Draw bounding boxes and category labels on the original image to visualize the detection results.

2. The intelligent detection method for steel surface defect identification according to claim 1, characterized in that: In step S31, the specific working process of the C3 module in the neck network is as follows: The C3 module adopts a SiLU structure, which includes a main branch and an auxiliary branch. The main branch includes a convolution model and a Bottleneck block model, and the auxiliary branch includes a convolution model. The output of the auxiliary branch is connected with the result calculated by the main branch, and then the feature information is further integrated through a convolution model. The three convolution models all perform 1×1 convolution operations.

3. The intelligent detection method for steel surface defect identification according to claim 2, characterized in that: In step S31, the specific working process of the Bottleneck block model of the C3 module in the neck network is as follows: The Bottleneck block model contains two branches. The first branch first halves the number of channels of the feature map through 1×1 convolution, then extracts features through 3×3 convolution and doubles the number of channels. The second branch performs residual connection through shortcut, adds the original input to the feature map after convolution operation, and realizes feature fusion.

4. The intelligent detection method for steel surface defect identification according to claim 3, characterized in that: In step S31, the specific working process of the bidirectional feature pyramid network in the neck network is as follows: First, remove the input nodes that have low contribution to feature fusion; Secondly, a bidirectional feature pyramid network locates and recognizes objects at various scales by introducing cross-scale connections between features from various scales; Third, a weighted feature fusion mechanism is introduced to fuse the features of each scale after connection. The features at each level are assigned corresponding weights, and these weights are automatically learned during the training process to ensure the best integrated features.

5. The intelligent detection method for steel surface defect identification according to claim 4, characterized in that: In step S32, when the target detection model is trained on the open source NEU surface defect detection dataset, the target detection model optimizes parameters by minimizing a loss function, where the loss function includes classification loss, bounding box regression loss, and confidence loss.

6. The intelligent detection method for steel surface defect identification according to claim 5, characterized in that: The method further includes step S5, defect type warning: storing the detection results and sending warnings for defects that reach a threshold.

7. The intelligent detection method for steel surface defect identification according to claim 6, characterized in that: It also includes step S6, process control: the controller determines whether the image sensor reaches the movement threshold. If not, it moves a specified distance and continues detection; if it reaches, it requests a reset instruction; if reset is selected, it continues with the new steel sample detection; if not reset is selected, the detection process ends.

8. An intelligent detection system for identifying surface defects in steel, comprising a computer program, characterized in that: The computer program can execute the intelligent detection method for identifying steel surface defects as described in any one of claims 1 to 7.

Images