Martensitic stainless steel strip production on-line monitoring system based on image recognition

Abstract

The present application relates to the technical field of image online identification, and especially relates to a martensitic stainless steel strip production online monitoring system based on image identification. The present application collects infrared images in the high-temperature discharging stage of hot forging blanks through an image acquisition module; extracts suspected defect contour lines through a feature extraction module; analyzes the presentation state of the contour and determines whether to start targeted augmentation through an image analysis module; divides the augmentation category and demarcates the target target area to execute image augmentation according to the distribution characteristics of the defect contour through a targeted processing module, and generates a local augmented image; extracts surface defect features and matches them with preset sample features through a defect identification module to evaluate the defect risk level. The present application adaptively starts image augmentation by analyzing the presentation state of the suspected defect contour, and enhances the distribution characteristics, effectively overcoming the problem that the acquisition of defect image information is limited due to the interference of the oxide skin, and improving the monitoring accuracy.

Description

Technical Field

[0001] This invention relates to the field of online image recognition technology, and in particular to an online monitoring system for the production of martensitic stainless steel strip based on image recognition. Background Technology

[0002] Hot forging is a crucial process in the forming of martensitic stainless steel parts. However, due to the poor thermal conductivity and sensitivity to surface cracks in martensitic stainless steel, it is highly susceptible to surface and near-surface defects such as microcracks, folds, and subcutaneous delamination during heating and forging. Failure to detect these defects in a timely manner during hot forging can significantly increase subsequent processing costs and even lead to product failure risks during service. Martensitic stainless steel hot-forged billets are typically covered with a thick oxide scale during the high-temperature unloading stage. Current technologies do not adequately consider how this oxide scale can mask microcracks, folds, and subcutaneous delamination defects on the surface of martensitic stainless steel hot-forged billets. Therefore, effective and rapid defect detection within the critical temperature range after hot forging and before the slow cooling process is of great significance for ensuring product quality and reducing scrap rates.

[0003] Existing technologies do not consider the interference of oxide scale in the identification of microcracks, folds, and subcutaneous delamination defects on the surface of hot-forged martensitic stainless steel billets. As a result, the acquisition of defect image information is affected by oxide scale, which affects the accuracy of online production monitoring. Summary of the Invention

[0004] To address this issue, the present invention provides an online monitoring system for martensitic stainless steel strip production based on image recognition. This system overcomes the problem in existing technologies that do not consider the interference of oxide scale in the identification of microcracks, folds, and subcutaneous delamination defects on the surface of hot-forged martensitic stainless steel billets, which leads to the acquisition of defect image information being affected by oxide scale interference and thus impacting the accuracy of online production monitoring.

[0005] To achieve the above objectives, the present invention provides an online monitoring system for martensitic stainless steel strip production based on image recognition, comprising: The image acquisition module is used to acquire infrared images of hot forging billets during the high-temperature discharge stage of hot forging billets of martensitic stainless steel. A feature extraction module, which is connected to the image acquisition module, is used to extract the outline of suspected defects from the infrared image; An image analysis module, connected to the feature extraction module, is used to analyze the presentation state of suspected defect contours and determine whether to perform targeted augmentation on the infrared image based on the presentation state. The targeting processing module is connected to the feature extraction module and the image analysis module respectively. It is used to determine the augmentation category based on the distribution characteristics of the suspected defect contour on the hot forging billet, determine the target target area in the infrared image for different augmentation categories, and perform image augmentation on the target target area to generate a local augmented image. The defect identification module, which is connected to the targeting processing module, is used to extract surface defect features in the locally augmented image and assess the defect risk level based on the matching results of the surface defect features and preset sample features.

[0006] Furthermore, the image acquisition module is used to acquire infrared images of each surface of the hot forging billet during the time interval between the final forging and the slow cooling process.

[0007] Furthermore, the feature extraction module is used to extract the outline of suspected defects, wherein, The feature extraction module calculates the neighborhood brightness gradient based on the grayscale brightness of each pixel in the infrared image and identifies pixels with brightness abrupt changes as candidate edge pixels. The feature extraction module connects several candidate edge pixels in sequence to form a suspected defect contour line.

[0008] Furthermore, the image analysis module is used to analyze the presentation state of suspected defect contours, wherein, The image analysis module is used to count the total length of all suspected defect contours and the cumulative length of closed contours, and to calculate the ratio of the cumulative length of closed contours to the total length of the contours.

[0009] Furthermore, the image analysis module is used to determine whether to perform targeted augmentation on the infrared image based on the presentation state, wherein, The image analysis module is used to compare the length ratio with a preset length ratio reference value; If the length ratio is less than or equal to the length ratio reference value, then targeted augmentation of the infrared image is determined.

[0010] Furthermore, the targeting processing module is used to determine the distribution characteristics of suspected defect contours on the hot-forged billet, wherein, The targeting processing module extracts the coordinates of pixels on the suspected defect contour line, calculates the average distance between the coordinates of pixels on the suspected defect contour line and the edge of the hot forging billet, and determines the average distance value as the distribution characteristic value of the suspected defect contour.

[0011] Furthermore, the targeting processing module is used to determine the augmentation category based on the comparison result between the distribution feature value and the preset feature discrimination value, wherein, If the distribution feature value is less than or equal to the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the edge augmentation category; If the distribution feature value is greater than the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the non-edge augmentation category.

[0012] Furthermore, the targeting processing module is used to define target regions for different augmentation categories, wherein, If the augmentation category is edge augmentation, the targeting processing module will define the continuous edge region as the target targeting region; If the augmentation category is a non-edge augmentation category, the targeting processing module will use the remaining image region after removing continuous edge regions as the target region.

[0013] The continuous edge region is a region formed by extending inward by a preset pixel distance based on each edge of the hot forging billet.

[0014] Furthermore, the targeting processing module is used to perform image augmentation on the target region to generate a locally augmented image, wherein, The targeting processing module establishes a pixel coordinate system with the center of the target region as the reference, and performs horizontal and vertical flipping and multi-angle rotation transformations on the target region to generate several local augmented images.

[0015] Furthermore, the defect identification module assesses the defect risk level based on the matching results between surface defect features and preset sample features, wherein, The defect identification module calculates the cosine similarity between the augmented explicit feature set and the pre-constructed feature set of standard defect samples of martensitic stainless steel item by item, and takes the average of all similarities as the comprehensive feature matching degree. The size of the defect is identified based on the comprehensive feature matching degree, and the risk level of the defect is assessed based on the size of the defect. The augmented explicit feature set includes a subset of gray-level difference features and a subset of defect geometric morphology features.

[0016] The beneficial effects of the technical solution presented in this application include: acquiring infrared images during the high-temperature unloading stage of hot forging billets through an image acquisition module, and extracting suspected defect contours through a feature extraction module, effectively capturing abnormal temperature gradient areas caused by internal material discontinuities such as subcutaneous delamination and microcracks; the image analysis module, by analyzing the presentation state of suspected defect contours, promptly detects insufficient defect information due to oxide scale obstruction, and adaptively initiates targeted augmentation processing when defect information is insufficient, thereby reducing the missed detection rate of defects such as microcracks and subcutaneous delamination; simultaneously, the targeted processing module delineates different target regions for edges and interiors respectively, achieving differentiated local enhancement and ensuring the effective extraction of weak defect information in infrared images; the defect recognition module extracts multi-dimensional surface defect features such as grayscale texture and defect geometry from the augmented image, and performs cosine similarity matching with standard sample features, effectively supporting online assessment and risk warning of the surface quality of martensitic stainless steel hot forging billets, overcoming the problems of insufficient consideration of oxide scale interference leading to limited acquisition of defect image information and insufficient accuracy of online production monitoring in existing technologies.

[0017] Furthermore, the present invention uses the degree of brightness change between adjacent pixels as the basis for determining the defect boundary, which can effectively distinguish between the gradual brightness change caused by uneven oxide thickness and the brightness change caused by discontinuity of the subsurface material. Even when the oxide scale is fully covered on the surface of the high-temperature blank, it can still accurately locate the boundary position of the real defect.

[0018] Furthermore, this invention uses contour morphology statistics to intuitively reflect the degree of occlusion of defect signals by oxide scale and the signal-to-noise ratio of the entire image. The ratio directly corresponds to the integrity and recognizability of the defect signal, thus enabling rapid evaluation of infrared image quality.

[0019] Furthermore, this invention only activates augmentation enhancement in scenarios where defect signals are broken and the risk of missed detection is high. On the one hand, it avoids the problem of global augmentation amplifying the oxide scale noise of the normal matrix, which would reduce the image signal-to-noise ratio. On the other hand, it significantly reduces the computational load per billet, ensuring the detection efficiency of continuous output from the hot forging production line.

[0020] Furthermore, for edge augmentation categories, the present invention delineates a narrow, continuous region along the edge to precisely match the banded distribution characteristics of edge defects, while avoiding a large area of ​​normal matrix in the center of the forging billet, thus minimizing the introduction of invalid background noise. For non-edge augmentation categories, after removing edge interference zones, the internal matrix region is locked to eliminate edge interference. By precisely limiting the augmentation range, it ensures that areas with high defect incidence are completely covered, while avoiding amplification of oxide scale noise throughout the augmentation process.

[0021] Furthermore, by flipping and rotating, this invention simulates the posture differences caused by the deviation of billet placement and the randomness of defect orientation in real production, enriching the feature presentation dimensions of weak defects and improving the robustness and scene adaptability of subsequent feature matching. On the other hand, it strictly constrains the augmentation operation to the local target area, and will not amplify the oxide noise of the whole image, effectively avoiding the drawback of the decrease in signal-to-noise ratio caused by global augmentation.

[0022] Furthermore, this invention employs a multi-dimensional feature system of grayscale texture and geometric shape to construct an identification system, and combines cosine similarity to achieve quantitative matching of defect features, avoiding the one-sidedness of single feature identification and improving the accuracy of defect type and size identification; based on the matching results, cracking risk levels are divided, providing a practical basis for online quality control of martensitic stainless steel hot forging billets. Attached Figure Description

[0023] Figure 1 This is a system block diagram of an online monitoring system for martensitic stainless steel strip production based on image recognition, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the target regions corresponding to each augmentation category in the embodiments of the present invention; In the image: 1 - continuous edge region, 2 - the image region remaining after removing the continuous edge region. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] It should be noted that in the description of this invention, the terms "upper," "lower," "inner," "outer," etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0027] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] With the upgrading of the precision hot forging industry and the rapid iteration of machine vision infrared imaging detection technology, martensitic stainless steel forging production lines using infrared vision online defect monitoring are gradually becoming more widespread. Despite continuous iterations in detection hardware, most current defect detection solutions for hot-forged martensitic stainless steel billets are developed for room-temperature workpieces without oxide scale coverage. Therefore, dedicated image recognition algorithms adapted to high-temperature oxide scale obscuring conditions are needed to fully utilize the detection capabilities of infrared online monitoring equipment in the hot forging unloading process. Existing image recognition-based metal surface defect detection methods are mostly geared towards room-temperature machined parts, employing global grayscale threshold segmentation and basic edge extraction strategies. These methods fail to consider the grayscale gradient interference caused by the full coverage of oxide scale on the high-temperature forging billet surface. This results in weak defect contours such as microcracks, folds, and subsurface delamination being submerged by oxide scale signals during the high-temperature unloading stage, leading to frequent missed and false detections during online identification, failing to meet the real-time quality inspection standards before the slow cooling of the hot forging process. Furthermore, most existing defect recognition training datasets are collected from room-temperature samples without oxide scale, failing to simulate the real-world conditions of random oxide scale thickness distribution in high-temperature forging billets (950℃~1150℃). This severely limits the model's generalization ability for defect recognition in industrial hot forging environments. In addition, existing recognition frameworks lack differentiated image enhancement mechanisms for edge stress concentration defects and subsurface delamination defects within the matrix. Global uniform augmentation tends to amplify oxide scale noise, resulting in insufficient precision in defect feature extraction and making it difficult to meet the high-precision online quality control requirements of martensitic stainless steel hot forging production.

[0029] Please see Figure 1 The diagram shown is a system block diagram of an online monitoring system for martensitic stainless steel strip production based on image recognition, according to an embodiment of the present invention. The online monitoring system for martensitic stainless steel strip production based on image recognition of the present invention includes: The image acquisition module is used to acquire infrared images of hot forging billets during the high-temperature discharge stage of hot forging billets of martensitic stainless steel. In this invention, the target acquisition station is the area before the martensitic stainless steel hot forging billet enters the slow cooling furnace after the final forging process. The surface of the billet in this area is stably maintained at a high temperature of 950℃~1150℃. This module uses multiple sets of short-wave infrared industrial cameras arranged in a ring as image acquisition hardware.

[0030] A feature extraction module, which is connected to the image acquisition module, is used to extract the outline of suspected defects from the infrared image; An image analysis module, connected to the feature extraction module, is used to analyze the presentation state of suspected defect contours and determine whether to perform targeted augmentation on the infrared image based on the presentation state. The targeting processing module is connected to the feature extraction module and the image analysis module respectively. It is used to determine the augmentation category based on the distribution characteristics of the suspected defect contour on the hot forging billet, determine the target target area in the infrared image for different augmentation categories, and perform image augmentation on the target target area to generate a local augmented image. The defect identification module, which is connected to the targeting processing module, is used to extract surface defect features in the locally augmented image and assess the defect risk level based on the matching results of the surface defect features and preset sample features.

[0031] In this invention, the feature extraction module, image analysis module, target processing module, and defect recognition module all belong to the data processing module category. The internal hardware structure of each of the above data processing modules is not limited. The module body or each internal computing unit can be constructed using logic components such as field-programmable logic devices, microprocessors, and computer processors, which will not be elaborated here.

[0032] Specifically, the image acquisition module is used to acquire infrared images of each surface of the hot forging billet during the time interval between the final forging and the slow cooling process.

[0033] In this invention, the temperature range for acquisition is the high-temperature range of 950℃ to 1150℃ after the hot forging billet is finally forged and discharged. Imaging is performed using a short-wave infrared imaging device arranged in a ring. Multiple camera units are evenly arranged along the circumference of the conveyor roller, corresponding to the circumferential side and end faces of the hot forging billet, respectively. The entire surface of the continuously moving hot forging billet is scanned and acquired synchronously to obtain infrared images of the hot forging billet with an oxide layer on the surface.

[0034] Specifically, the feature extraction module is used to extract the outline of suspected defects, wherein, The feature extraction module calculates the neighborhood brightness gradient based on the grayscale brightness of each pixel in the infrared image and identifies pixels with brightness abrupt changes as candidate edge pixels. The feature extraction module connects several candidate edge pixels in sequence to form a suspected defect contour line.

[0035] In this invention, the neighborhood brightness gradient uses a 3×3 pixel calculation window to solve the gray-scale brightness difference in the horizontal and vertical directions, and synthesizes the gradient amplitude of each pixel. The criterion for determining the brightness abrupt change feature is that the gradient amplitude of the pixel reaches a preset gradient threshold. The preset gradient threshold is based on the gradient mean of the normal substrate area of ​​the forging billet. The intact substrate area covered with normal oxide scale has a gradual change in neighborhood brightness due to the gradual change in the thickness of the oxide scale. The average value of the synthesized gradient amplitude obtained by calculation is usually 8 to 12. The threshold for determining the defect abrupt change is 2 to 3 times the average gradient of the substrate. The value range of the preset gradient threshold is 20 to 30. Preferably, in this invention, the preset gradient threshold is 25.

[0036] In this calculation, a 3×3 pixel neighborhood is taken centered on the current pixel. The horizontal and vertical gray-level difference components corresponding to the pixel are obtained through the Sobel operator. Then, the square root of the sum of the squares of the two directional difference components is calculated to obtain the gradient magnitude of the pixel. This embodiment uses a standardized Sobel convolution kernel, with the horizontal convolution kernel matrix being [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] and the vertical convolution kernel matrix being [[-1, -2, -1], [0, 0, 0], [1, 2, 1]]. Before the convolution operation, the gray-level values ​​of the entire infrared image are normalized and mapped to the range of 0 to 255. The Sobel operator is a commonly used processing method in the field of visual image processing, and its specific calculation method is well known and will not be elaborated here.

[0037] In this invention, candidate pixels that are spatially adjacent with a pixel spacing of no more than 3 pixel units are sequentially connected to form a continuous contour line.

[0038] It is understandable that martensitic stainless steel hot-forged billets are covered with a randomly distributed oxide scale layer of varying thickness at high temperatures. Areas with thick oxide scale have lower overall grayscale brightness, while areas with thin oxide scale have higher overall grayscale brightness. Subsurface cracks, delamination, and other real defects are essentially due to discontinuities in the internal material, causing abrupt changes in heat conduction at the defect boundaries. This corresponds to abrupt changes in brightness between adjacent pixels in infrared images—a brightness gradient significantly higher than the brightness difference created by the gradual change in oxide scale thickness. Therefore, using neighborhood brightness gradients to filter pixels with abrupt brightness changes can effectively filter out interference from the overall brightness gradient caused by uneven oxide scale thickness, accurately pinpointing the boundary location of real defects.

[0039] Specifically, the image analysis module is used to analyze the presentation state of suspected defect contours, wherein, The image analysis module is used to count the total length of all suspected defect contours and the cumulative length of closed contours, and to calculate the ratio of the cumulative length of closed contours to the total length of the contours.

[0040] Those skilled in the art should understand that, under the occlusion effect of oxide scale, the true and complete defect contour will be cut into multiple discontinuous and fragmented non-closed contour segments by the oxide scale; only in areas with thin oxide scale, large defect depth, and strong temperature gradient signal can the defect contour maintain a continuous and closed complete shape. Therefore, the ratio of the cumulative length of closed contours to the total length of all contours can directly quantify the integrity and signal-to-noise ratio of the defect signal in the entire image: the higher the ratio, the higher the proportion of clearly identifiable defects in the image, and the smaller the impact of oxide scale occlusion; the lower the ratio, the more severe the fragmentation of the defect contour, the more weak signal defects caused by oxide scale occlusion, and the higher the risk of missed detection in direct identification.

[0041] In practice, the contour length is calculated by accumulating the pixel Euclidean distances, and the length of a single contour line is the sum of the pixel distances between adjacent candidate pixels.

[0042] Specifically, the image analysis module is used to determine whether to perform targeted augmentation on the infrared image based on the presentation state, wherein, The image analysis module is used to compare the length ratio with a preset length ratio reference value; If the length ratio is greater than the length ratio reference value, then it is determined that the infrared image will not be targeted augmented; If the length ratio is less than or equal to the length ratio reference value, then targeted augmentation of the infrared image is determined.

[0043] In this invention, the length ratio reference value is obtained through pre-test calibration: At least 30 samples of the same type of martensitic stainless steel hot-forged billets are selected, and infrared images are collected under different oxide scale thicknesses. The length ratio corresponding to a satisfactory level of manually annotated defect identification is used as a benchmark to determine the length ratio reference value. When the length ratio reference value is lower than this reference value, it indicates that the degree of contour fragmentation caused by oxide scale obstruction is sufficient to cause missed detection, requiring targeted augmentation processing. For example, the length ratio reference value can be set to 0.4.

[0044] In this invention, if global image augmentation is uniformly applied to all hot-forged billet images, it will amplify the oxide scale noise in the normal matrix area, thus reducing the signal-to-noise ratio for defect identification. Furthermore, it will significantly increase the processing power of a single frame, failing to meet the online detection timeliness requirements of continuous hot-forging production lines. By activating augmentation only when the defect signal integrity is insufficient and the risk of missed detection increases, the identification accuracy in weak defect scenarios can be guaranteed, while maintaining detection efficiency in high-quality image scenarios.

[0045] Specifically, the targeting module is used to determine the distribution characteristics of suspected defect contours on the hot-forged billet, wherein, The targeting processing module extracts the coordinates of pixels on the suspected defect contour line, calculates the average distance between the coordinates of pixels on the suspected defect contour line and the edge of the hot forging billet, and determines the average distance value as the distribution characteristic value of the suspected defect contour.

[0046] In this invention, the edges of the hot forging billet are obtained by global contour extraction from an infrared image: grayscale thresholding is performed on the original infrared image to extract the overall outer contour boundary of the hot forging billet, and the corner edges of the outer contour are the edges of the hot forging billet.

[0047] Specifically, the targeting processing module is used to determine the augmentation category based on the comparison result between the distribution feature value and the preset feature discrimination value, wherein, If the distribution feature value is less than or equal to the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the edge augmentation category; If the distribution feature value is greater than the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the non-edge augmentation category.

[0048] In this invention, the feature differentiation value corresponds to the concentrated influence zone of the superposition of forging stress and phase transformation stress at the edge of the hot forged martensitic stainless steel billet. The physical width is about 10% to 15% of the minimum side length of the billet. This value is then converted into pixel distance to determine the final feature differentiation value.

[0049] Understandably, the edges and corners of hot-forged martensitic stainless steel billets are areas of concentrated forging deformation stress, and the superposition effect of phase transformation stress during cooling is significant, making them high-risk areas for microcracks. Furthermore, oxide scale tends to accumulate and detach unevenly in these areas, further exacerbating defect signal obscuring. In contrast, the stress level in the matrix region inside the billet is relatively gentle, and defects are mostly deep subsurface layers, resulting in weak but concentrated signals. Therefore, classifying the augmentation scenarios into two categories based on the relative distribution of defects to the edges allows for differentiated augmentation strategies tailored to the defect causes and signal characteristics in different regions.

[0050] Specifically, the targeting processing module is used to define target regions for different augmentation categories, wherein, If the augmentation category is edge augmentation, the targeting processing module will define the continuous edge region as the target targeting region; If the augmentation category is a non-edge augmentation category, the targeting processing module will use the remaining image region after removing continuous edge regions as the target targeting region; The continuous edge region is a region formed by extending inward by a preset pixel distance based on each edge of the hot forging billet.

[0051] Please see Figure 2As shown, it is a schematic diagram of the target target area corresponding to each augmentation category in the embodiment of the present invention. If the augmentation category is the edge augmentation category, the targeting processing module defines the continuous edge region 1 as the target target area. The continuous edge region is the area formed by extending a preset pixel distance a inward from each edge of the hot forging billet as a reference. If the augmentation category is the non-edge augmentation category, the targeting processing module takes the image area 2 remaining after removing the continuous edge region 1 as the target target area.

[0052] In this invention, the preset pixel distance is converted into physical width, which is about 10% to 15% of the minimum side length of the blank. For example, the preset pixel distance is converted into physical width as 10% of the minimum side length of the blank.

[0053] In this invention, for edge augmentation categories, defects are concentrated in narrow stress bands near the edges. Therefore, a continuous edge region extending inward along the edge is used as the target range to accurately cover defect-prone areas while avoiding large normal areas in the center of the forging billet, thus preventing the introduction of invalid background noise. For non-edge augmentation categories, defects are mostly distributed in the center of the matrix. Eliminating edge regions can remove interference, allowing the augmentation operation to focus on weak defect signals in the non-edge internal matrix, ensuring the accuracy of feature extraction after augmentation.

[0054] Specifically, the targeting processing module is used to perform image augmentation on the target region to generate a locally augmented image, wherein, The targeting processing module establishes a pixel coordinate system with the center of the target region as the reference, and performs horizontal and vertical flipping and multi-angle rotation transformations on the target region to generate several local augmented images.

[0055] In this invention, the flipping operation includes horizontal flipping and vertical flipping, and the rotation operation takes the geometric center of the target area as the origin and selects multiple uniformly spaced rotation angles within a small angle range for rotation transformation; the flipping and rotation operations only act on pixels within the target area, while the image outside the area remains unchanged, ultimately generating multiple locally augmented images containing different posture defect features.

[0056] For example, the rotation angle is selected from -15° to +15°, and samples are uniformly taken at 5° intervals to generate a total of 7 sets of rotated images. After superimposing the original image of the target area with horizontal and vertical flips, a single target area original image can generate 21 independent local augmented images.

[0057] In this invention, the billet placement angle has random deviations, and the same type of defects will present different directions and postures in infrared images. By performing multi-axis flipping and angle rotation on the target area, the different spatial postures of defects in real production can be simulated, enriching the feature dimensions of weak defects and improving the robustness of subsequent feature matching. At the same time, this operation is only performed within the defined target area and will not amplify the oxide scale noise in the entire domain. While enhancing the defect features, the image signal-to-noise ratio is maintained, which is suitable for the weak defect enhancement requirements of high-temperature infrared images.

[0058] Specifically, the defect identification module assesses the defect risk level based on the matching results between surface defect features and preset sample features, wherein, The defect identification module calculates the cosine similarity between the augmented explicit feature set and the pre-constructed feature set of standard defect samples of martensitic stainless steel item by item, and takes the average of all similarities as the comprehensive feature matching degree. The size of the defect is identified based on the comprehensive feature matching degree, and the risk level of the defect is assessed based on the size of the defect. The augmented explicit feature set includes a subset of gray-level difference features and a subset of defect geometric morphology features.

[0059] In this invention, the grayscale difference feature subset includes the average grayscale difference between the defect area and the surrounding normal substrate area within the target region, used to characterize the degree of light and dark distinction between the defect and the substrate; the defect geometric shape feature subset includes the total length of the defect contour and the area enclosed by the contour, used to characterize the overall scale of the defect. When calculating the cosine similarity, the grayscale difference feature, the geometric shape feature, and the corresponding feature vectors in the standard defect sample feature set are compared, and the average of the similarity results of the two dimensions is taken to obtain the comprehensive feature matching degree.

[0060] The pre-constructed standard defect sample feature set for martensitic stainless steel includes three types of defect samples: microcracks, folds, and subcutaneous delamination. Each type of sample covers different defect sizes from 0.1 mm to 3 mm. The feature vectors are normalized to the [0, 1] interval before cosine similarity is calculated to eliminate matching bias caused by differences in grayscale and size magnitude. The cosine similarity calculation result ranges from 0 to 1. The closer the value is to 1, the higher the defect matching degree. The cosine similarity calculation formula is well-known and will not be elaborated here.

[0061] This document provides an example of pre-constructing a standard defect sample feature set for martensitic stainless steel: Multiple batches of hot-forged martensitic stainless steel billets are selected, and microcracks, folds, and subcutaneous delamination defects of different depths and widths are generated through forging and artificial prefabrication. All prefabricated defects are confirmed to have their true dimensions by metallographic sectioning. Simultaneously, intact billets of the same specifications without any surface defects are prepared. The defective and intact billets are placed at a high-temperature station (950℃~1150℃) identical to the production line, and infrared grayscale images are captured using the same type of infrared camera with the same resolution and acquisition parameters. Quality inspectors select all actual defect areas in the infrared images and record the actual physical dimensions of each defect. The labeled sample infrared images are then processed... The same gradient calculation and contour connectivity algorithm as the feature extraction module of this invention are used to uniformly extract the average gray-level difference, total contour length, and contour-enclosed area of ​​the defect region, generating a single set of standard feature vectors. At least 100 independent sample images are collected for each type of defect, corresponding to the generation of 100 sets of independent feature vectors. At the same time, at least 200 sets of feature vectors of intact billet samples are recorded. The gray-level difference, length, and area features of all defective and intact billet samples are mapped and normalized to the [0, 1] value range. They are classified and stored according to defect type to form a standard defect sample feature set that can be directly called, ensuring that the augmented explicit features extracted by the online detection end are completely consistent with the feature dimensions and units in the library, and that the cosine similarity calculation is comparable.

[0062] The mapping normalization to the [0, 1] numerical range can be achieved using minimum and maximum normalization methods: for single-class features such as gray-level difference, contour length, and contour area, the maximum value X of the feature is obtained by traversing all samples. max Minimum value X min , using the formula X1=(XX min ) / (X max -X min The feature value X is mapped to the interval 0 to 1 to obtain the specific mapped value X1.

[0063] The defect risk level is divided into three levels: low, medium, and high, based on the comprehensive feature matching degree and defect size parameters. Low defect risk corresponds to small defect size, low matching degree, and no obvious cracking risk; medium defect risk corresponds to the presence of surface defects; and high defect risk corresponds to the presence of large-sized deep defects with a significant cracking risk. For example, this embodiment uses the following quantitative judgment thresholds: comprehensive feature matching degree < 0.4 and equivalent defect size < 0.3 mm is judged as low defect risk; 0.4 ≤ comprehensive feature matching degree ≤ 0.7 and 0.3 mm ≤ equivalent defect size ≤ 1.0 mm is judged as medium defect risk; and comprehensive feature matching degree > 0.7 or equivalent defect size > 1.0 mm is judged as high defect risk. The equivalent defect size is obtained by converting the area enclosed by the defect contour into the equivalent circular diameter, with the unit uniformly in millimeters. The system stores the matching degree, defect size, risk level, and collection time of each blank in real time, forming a traceable production quality database.

[0064] The defect risk level results output by this invention can be directly connected to the slow cooling furnace control system of the hot forging production line. Different slow cooling holding time and stepped cooling rate are matched for billets with different risk levels, forming a complete closed loop from the defect detection end to the process control end. While detecting defects, the billet cooling cracking scrap rate is reduced simultaneously, realizing online quality control of the entire process of hot forging production of martensitic stainless steel.

[0065] This invention uses the Modbus TCP industrial communication protocol to interact in real time with the production line PLC and slow cooling furnace controller. Hot forging billets with low defect risk are subjected to standard slow cooling process with a holding time of 60 minutes and a stepped cooling rate of 5℃ / min. Hot forging billets with medium defect risk have an extended holding time of 90 minutes and a cooling rate of 3℃ / min. Hot forging billets with high defect risk are diverted to the re-inspection station, where defects are manually removed before entering the slow cooling process. At the same time, an audible and visual alarm is pushed to the production line central control terminal. All process control parameters are archived synchronously with the billet inspection data.

[0066] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the 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. An online monitoring system for martensitic stainless steel strip production based on image recognition, characterized in that, include: The image acquisition module is used to acquire infrared images of hot forging billets during the high-temperature discharge stage of hot forging billets of martensitic stainless steel. A feature extraction module, which is connected to the image acquisition module, is used to extract the outline of suspected defects from the infrared image; An image analysis module, connected to the feature extraction module, is used to analyze the presentation state of suspected defect contours and determine whether to perform targeted augmentation on the infrared image based on the presentation state. The targeting processing module is connected to the feature extraction module and the image analysis module respectively. It is used to determine the augmentation category based on the distribution characteristics of the suspected defect contour on the hot forging billet, determine the target target area in the infrared image for different augmentation categories, and perform image augmentation on the target target area to generate a local augmented image. The defect identification module, which is connected to the targeting processing module, is used to extract surface defect features in the locally augmented image and assess the defect risk level based on the matching results of the surface defect features and preset sample features.

2. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 1, characterized in that, The image acquisition module is used to acquire infrared images of each surface of the hot forging billet during the time interval between the final forging and the slow cooling process.

3. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 1, characterized in that, The feature extraction module is used to extract the outline of suspected defects, wherein, The feature extraction module calculates the neighborhood brightness gradient based on the grayscale brightness of each pixel in the infrared image and identifies pixels with brightness abrupt changes as candidate edge pixels. The feature extraction module connects several candidate edge pixels in sequence to form a suspected defect contour line.

4. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 3, characterized in that, The image analysis module is used to analyze the presentation status of suspected defect contours, wherein, The image analysis module is used to count the total length of all suspected defect contours and the cumulative length of closed contours, and to calculate the ratio of the cumulative length of closed contours to the total length of the contours.

5. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 4, characterized in that, The image analysis module is used to determine whether to perform targeted augmentation on the infrared image based on the presentation state, wherein... The image analysis module is used to compare the length ratio with a preset length ratio reference value; If the length ratio is less than or equal to the length ratio reference value, then targeted augmentation of the infrared image is determined.

6. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 1, characterized in that, The targeting module is used to determine the distribution characteristics of suspected defect contours on the hot-forged billet, wherein... The targeting processing module extracts the coordinates of pixels on the suspected defect contour line, calculates the average distance between the coordinates of pixels on the suspected defect contour line and the edge of the hot forging billet, and determines the average distance value as the distribution characteristic value of the suspected defect contour.

7. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 6, characterized in that, The targeting processing module is used to determine the augmentation category based on the comparison result between the distribution feature value and the preset feature discrimination value, wherein... If the distribution feature value is less than or equal to the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the edge augmentation category; If the distribution feature value is greater than the feature discrimination value, the targeting processing module determines the augmentation category of the suspected defect contour as the non-edge augmentation category.

8. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 7, characterized in that, The targeting processing module is used to define target regions for different augmentation categories, wherein, If the augmentation category is edge augmentation, the targeting processing module will define the continuous edge region as the target targeting region; If the augmentation category is a non-edge augmentation category, the targeting processing module will use the remaining image region after removing continuous edge regions as the target targeting region; The continuous edge region is a region formed by extending inward by a preset pixel distance based on each edge of the hot forging billet.

9. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 8, characterized in that, The targeting processing module is used to perform image augmentation on the target region to generate a locally augmented image, wherein... The targeting processing module establishes a pixel coordinate system with the center of the target region as the reference, and performs horizontal and vertical flipping and multi-angle rotation transformations on the target region to generate several local augmented images.

10. The online monitoring system for martensitic stainless steel strip production based on image recognition according to claim 8, characterized in that, The defect identification module assesses the defect risk level based on the matching results between surface defect features and preset sample features, wherein... The defect identification module calculates the cosine similarity between the augmented explicit feature set and the pre-constructed feature set of standard defect samples of martensitic stainless steel item by item, and takes the average of all similarities as the comprehensive feature matching degree. The size of the defect is identified based on the comprehensive feature matching degree, and the risk level of the defect is assessed based on the size of the defect. The augmented explicit feature set includes a subset of gray-level difference features and a subset of defect geometric morphology features.

Images