A method for detecting debris in metal powder based on image recognition

Abstract

The present invention relates to the field of image data processing technology, and discloses a method for detecting impurities in metal powders based on image recognition, comprising: irradiating the metal powder with polarized light and synchronously collecting images of the orthogonal polarization components of its reflected light, and detecting impurities by calculating a differential image. The present invention utilizes the essential difference between the responses of metal powder and impurities to polarized light, causing qualified powder to spontaneously form a uniform dark field in the differential image, while impurities generate significant signals due to optical anisotropy, thereby breaking through the blind spot of traditional methods for detecting microscopic impurities. The method further multiplexes total light flux data for time series analysis to achieve synchronous recognition of impurity morphological characteristics, forming a dual-dimensional detection capability of composition and morphology. The system implements adaptive threshold adjustment based on the statistical characteristics of the differential image to ensure long-term operational stability.

Description

Technical Field

[0001] The invention relates to a method for detecting foreign matter in metal powder based on image recognition, and belongs to the technical field of image data processing. Background Art

[0002] Current mainstream detection methods rely on high-resolution imaging systems combined with complex AI algorithms to analyze morphological features. This type of technology faces three essential limitations: first, the light intensity contrast between microscopic impurities and the substrate tends to disappear. Existing methods are limited to a single dimension of light intensity information, making it difficult to distinguish foreign objects with similar materials; second, in the dynamic environment of high-speed production lines, powder flow causes motion blur and image registration distortion, and traditional multi-sensor solutions significantly increase system complexity; third, the fixed threshold judgment mechanism cannot adapt to variable drift such as light source aging and environmental disturbances, and requires frequent manual calibration.

[0003] The industry has attempted to optimize through multispectral imaging or deep learning, but due to high hardware costs or insufficient model generalization, it has failed to overcome the physical bottleneck of information capture. This is particularly true in additive manufacturing powder recycling scenarios, where the morphological grayscale of oxidized particles and the substrate are highly similar, making reliable identification difficult with existing technologies. A more fundamental contradiction lies in the fact that the existing paradigm overly relies on improving image clarity to increase algorithmic complexity, failing to explore the deeper physical properties of the interaction between matter and light. Therefore, how to achieve the essential differentiation of metal powder impurities based on image data processing technology, while simultaneously addressing the issues of dynamic interference and long-term stability, has become the technical problem addressed by this invention. Summary of the Invention

[0004] The present invention provides a method for detecting impurities in metal powder based on image recognition. Its main purpose is to solve the problems of existing metal powder impurity detection methods such as missed detection of microscopic impurities, poor adaptability to dynamic environments, and insufficient long-term operation stability due to the single information dimension.

[0005] To achieve the above objectives, the present invention provides a method for detecting impurities in metal powder based on image recognition, the method comprising the following steps:

[0006] Step a: constructing a polarized illumination field, wherein an industrial LED surface light source is used, and a circular polarizer is arranged in front of the industrial LED surface light source to irradiate the flowing metal powder with circularly polarized light;

[0007] Step b: synchronously capturing at least two orthogonal linear polarization component images of light reflected from the metal powder, wherein an industrial black-and-white camera is used, and a polarization beam splitter prism or a micro-polarization array filter is disposed in front of the lens of the industrial black-and-white camera, so that the camera can receive at least two orthogonal linear polarization component images in different areas of its single image sensor in a single exposure;

[0008] Step c, performing model-free differential information extraction, wherein each frame of image captured by the camera is processed, the image is divided into a first linear polarization component image and a second linear polarization component image, and a normalized differential image is calculated, where the intensity of each pixel of the normalized differential image is the absolute value of the intensity difference of the pixel in the first linear polarization component image and the second linear polarization component image, divided by the sum of the intensities of the two images and a non-zero positive number;

[0009] Step d: identifying pixel areas in the normalized difference image whose pixel intensity is higher than a preset determination threshold as impurities in the metal powder.

[0010] Preferably, in step d, the identification comprises: obtaining a pixel intensity histogram of the normalized difference image; performing statistical analysis on the pixel intensity histogram to determine the mean value representing the main distribution of qualified metal powder and standard deviation ; and based on the mean and standard deviation , dynamically generate the decision threshold, where the decision threshold , It is a preset coefficient with a value range of 5 to 6.

[0011] Preferably, the splitting ratio of the polarization beam splitter prism or the micro-polarization array filter is such that at least two orthogonal linear polarization component images obtain a balanced intensity distribution within the dynamic response range of the camera sensor.

[0012] Preferably, the normalized differential image calculation in step c is completed in real time on an embedded processor.

[0013] Preferably, it also includes: step e, obtaining a time-series image sequence composed of a total luminous flux image formed by superimposing the first linear polarization component image and the second linear polarization component image, wherein the total luminous flux image intensity = the first linear polarization component image intensity + the second linear polarization component image intensity; step f, in the time-series image sequence, using the optical flow method or correlation matching to perform micro-trajectory tracking on the moving particles to form a dynamic observation window that follows the movement of specific particles; and step g, analyzing the change characteristics of the pixel brightness value in the dynamic observation window with time, wherein the time standard deviation of the dynamic observation window on the time-series image sequence is calculated; and based on the time standard deviation, identifying impurities with non-spherical morphological characteristics in the metal powder.

[0014] Preferably, in step g, if an area presents an intensity higher than the judgment threshold on the normalized difference image, it is judged as a composition difference type impurity; if an area presents an intensity lower than the judgment threshold on the normalized difference image, but presents a time standard deviation higher than the preset morphological threshold in its corresponding total luminous flux time series analysis, it is judged as a morphological difference type impurity.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. Through the orthogonal polarization difference mechanism, the reflected light of the metal powder is decomposed into two orthogonal polarization components and a differential image is generated. When the qualified powder spontaneously forms a uniform dark field in the differential image due to its high reflection symmetry, the optical anisotropy of the impurity triggers a sudden change in pixel intensity. This mechanism achieves the physical self-elimination of background signals and the essential highlighting of impurity characteristics without relying on high-resolution imaging or complex algorithms, making traditionally difficult-to-distinguish microscopic foreign objects naturally identifiable at the image data level.

[0017] 2. The total luminous flux data stream from the differential image calculation process is reused and combined with time-series analysis to construct a dynamic observation window. When morphologically abnormal impurities such as metal shavings generate reflection intensity pulsations in the flow, their spatiotemporal fluctuation characteristics are decoupled and extracted through standard deviation calculation. This process forms an orthogonal information dimension with the polarization differential mechanism—composition differences are decoded by the polarization response, and morphological anomalies are captured by the time-varying characteristics of the luminous flux. The two work together to achieve dual verification of the physical properties of impurities, providing diagnostic-level data for production line process optimization that cannot be obtained with traditional single-point detection.

[0018] 3. A statistical self-anchoring mechanism based on the global histogram of the differential image transforms the main distribution of qualified powder into a system benchmark. By real-time tracking the mean drift and dispersion of the main peak of the histogram, a judgment threshold that aligns with the current operating conditions is dynamically generated. When ambient light changes or equipment aging cause signal baseline fluctuations, the system automatically maintains detection sensitivity based on statistical laws. This fundamentally avoids the risk of misjudgment caused by physical quantity drift in traditional fixed threshold solutions, and provides the detection logic with continuous and stable autonomous error correction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a module structure diagram of the impurity detection system based on polarization imaging and image analysis of the present invention;

[0020] Figure 2 Schematic diagram of the system structure of the polarized illumination and acquisition unit and the embedded processing unit of the present invention;

[0021] Figure 3 A schematic diagram of the timing logic of the adaptive update process of the determination threshold of the present invention;

[0022] Figure 4 This is a diagram showing the actual detection effect of the metal powder impurity detection system of the present invention.

[0023] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0024] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] The present invention provides a method for detecting foreign matter in metal powder based on image recognition, the method comprising the following steps:

[0026] Step a: constructing a polarized illumination field, wherein an industrial LED surface light source is used, and a circular polarizer is arranged in front of the industrial LED surface light source to irradiate the flowing metal powder with circularly polarized light;

[0027] Step b: synchronously capturing at least two orthogonal linear polarization component images of light reflected from the metal powder, wherein an industrial black-and-white camera is used, and a polarization beam splitter prism or a micro-polarization array filter is disposed in front of the lens of the industrial black-and-white camera, so that the camera can receive at least two orthogonal linear polarization component images in different areas of its single image sensor in a single exposure;

[0028] Step c, performing model-free differential information extraction, wherein each frame of image captured by the camera is processed, the image is divided into a first linear polarization component image and a second linear polarization component image, and a normalized differential image is calculated, where the intensity of each pixel of the normalized differential image is the absolute value of the intensity difference of the pixel in the first linear polarization component image and the second linear polarization component image, divided by the sum of the intensities of the two images and a non-zero positive number; specifically, each frame of image captured by the camera is processed, the image is divided into a first linear polarization component image and a second linear polarization component image, and a normalized differential image of the first linear polarization component image and the second linear polarization component image is calculated, where the normalized differential image intensity = |first linear polarization component image intensity - second linear polarization component image intensity| / (first linear polarization component image intensity + second linear polarization component image intensity + a non-zero positive number);

[0029] Step d: identifying pixel areas in the normalized difference image whose pixel intensity is higher than a preset determination threshold as impurities in the metal powder.

[0030] Preferably, in step d, the identification comprises: obtaining a pixel intensity histogram of the normalized difference image; performing statistical analysis on the pixel intensity histogram to determine the mean value representing the main distribution of qualified metal powder and standard deviation ; and based on the mean and standard deviation , dynamically generate the decision threshold, where the decision threshold , It is a preset coefficient with a value range of 5 to 6.

[0031] Preferably, the splitting ratio of the polarization beam splitter prism or the micro-polarization array filter is such that at least two orthogonal linear polarization component images obtain a balanced intensity distribution within the dynamic response range of the camera sensor.

[0032] Preferably, the normalized differential image calculation in step c is completed in real time on an embedded processor.

[0033] Preferably, it also includes: step e, obtaining a time-series image sequence composed of a total luminous flux image formed by superimposing the first linear polarization component image and the second linear polarization component image, wherein the total luminous flux image intensity = the first linear polarization component image intensity + the second linear polarization component image intensity; step f, in the time-series image sequence, using the optical flow method or correlation matching to perform micro-trajectory tracking on the moving particles to form a dynamic observation window that follows the movement of specific particles; and step g, analyzing the change characteristics of the pixel brightness value in the dynamic observation window with time, wherein the time standard deviation of the dynamic observation window on the time-series image sequence is calculated; and based on the time standard deviation, identifying impurities with non-spherical morphological characteristics in the metal powder.

[0034] Preferably, in step g, if an area presents an intensity higher than the judgment threshold on the normalized difference image, it is judged as a composition difference type impurity; if an area presents an intensity lower than the judgment threshold on the normalized difference image, but presents a time standard deviation higher than the preset morphological threshold in its corresponding total luminous flux time series analysis, it is judged as a morphological difference type impurity.

[0035] At the same time, in the actual implementation process, in order to ensure the stability of the normalized differential image calculation, the non-zero positive number used to prevent division by zero described in the present invention is preferably set to one to two times the minimum grayscale step value supported by the image sensor. For example, under 8-bit grayscale encoding conditions, the value range can be selected between 1 and 3 to offset the numerical fluctuations in weak signal areas. It can not only effectively avoid calculation anomalies with a denominator of zero, but also maintain the consistency of image contrast without introducing significant noise. At the same time, the differential image refers to the image result after the above-mentioned normalization processing; in order to ensure the robustness and engineering adaptability of impurity judgment, the coefficient range of the threshold dynamic adjustment strategy adopted in the present invention is between five and six. This value range is based on the empirical statistical results extracted from a large number of different batches of metal powder image samples. It can significantly improve the recognition accuracy of composition-differential impurities while ensuring a low false alarm rate. In actual applications, it can also be reasonably fine-tuned according to the powder material, particle size distribution or production environment to adapt to The detection sensitivity requirements under different working conditions and the dynamic response range of the image sensor refer to the grayscale intensity change range that the sensor can perceive and accurately output, which is usually limited by its bit depth and the linear range of the analog signal amplification circuit. Ensuring that the acquired image is within this range helps to maintain the consistency and comparability of the image signal; and in the powder particle tracking process, in response to complex situations such as particle occlusion, stacking or dense distribution that may occur during the flow process, the system introduces an anomaly detection strategy based on time continuity on the basis of the optical flow method or the correlation matching method, that is, target particles with smooth trajectory changes and continuous pixel brightness fluctuations are preferentially selected for observation within the tracking window. When short-term occlusion or intensity mutation occurs in the image area, it is eliminated or corrected through the inter-frame position prediction and brightness recovery mechanism to avoid misjudging briefly overlapping particles or ghosting as real impurity signals, thereby improving the recognition accuracy and stability during dynamic tracking. These are all extended implementation methods that are known to ordinary technicians in this field.

[0036] Example 1: The present invention proposes an image recognition method for detecting impurities in metal powders, which is based on the polarization light response difference and differential image calculation mechanism, combined with image statistical judgment and time series feature analysis, so as to achieve effective recognition and classification of impurities. In terms of the construction of the optical lighting system, the present invention adopts an industrial-grade LED surface light source, and configures a circular polarizer at its front end, so that the light irradiated to the surface of the metal powder is in a circularly polarized state. Here, the circular polarizer refers to an optical device that can convert linearly polarized light into circularly polarized light, with stable phase delay performance and directional consistency, and is suitable for achieving stable lighting in a dynamic powder flow field. Through this setting, the optical response difference of the object under test in different structural directions can be effectively enhanced, providing a basis for subsequent polarization component analysis; in the image acquisition stage, an industrial black and white camera equipped with a polarization beam splitter prism or a micro-polarization array filter is used to make The camera is able to simultaneously acquire two orthogonal linearly polarized component images of light reflected from metal powder in different spatial regions of its image sensor during a single exposure. The polarization beamsplitter prism separates the incident light according to its polarization direction based on the principle of optical birefringence, while the micro-polarization array filter constructs microstructure units with directional response on the image sensor to achieve spatial encoding of different polarization components. To ensure that the two orthogonal component images have a nearly consistent intensity distribution within the dynamic range, the prism or filter should have a balanced splitting ratio. This property is used in this method to improve the stability of the normalized difference image calculation.

[0037] The present invention implements a model-free differential information extraction mechanism in the image processing module. For each frame of image captured by the camera, the mechanism first completes the segmentation of two orthogonal polarization component images, and then calculates a normalized differential image. The pixel intensity of the differential image is defined as the absolute value of the intensity difference between the first polarization component image and the second polarization component image, divided by the sum of the two intensities and a non-zero positive number. The setting range of the non-zero positive number should be selected according to the minimum response accuracy of the image sensor to ensure that the differential result is not distorted by zero division. Through this calculation, the optical difference characteristics of the impurity area can be significantly highlighted, while suppressing the homogeneous signal of the qualified powder background, thereby obtaining a differential image with natural contrast; and in the recognition link, the system further performs a global pixel intensity histogram analysis on the normalized differential image, extracts the mean and standard deviation of the main distribution peak representing the qualified metal powder, and generates a dynamic judgment threshold based on this. The threshold is calculated as the main distribution mean plus the product of a preset coefficient and the standard deviation. The coefficient range is between five and six, which is used to exclude normal powder. The statistical fluctuations caused by this threshold are judged as debris areas based on the area where the intensity exceeds this value in the differential image. This process does not rely on model training or prior templates, and has good adaptability and environmental robustness. In addition, in order to further identify the morphological characteristics of impurities, the present invention introduces a time-series image analysis mechanism. Specifically, the superimposed image of the two polarization component images is reused to construct a total luminous flux image and form a time-series image sequence. By tracking the tiny motion trajectories of the powder particles in the image sequence, a dynamic observation window is established. The particle tracking process can be implemented based on optical flow method or regional correlation matching technology to ensure that the observation window always locks on the same target particle. Within the observation window, the pixel brightness value in the time-series image is statistically analyzed to calculate its time standard deviation value. This standard deviation value is used to reflect the degree of fluctuation of the reflection intensity of the particles during the flow process. Generally speaking, qualified powder has a low standard deviation due to its regular shape and stable reflection; while impurities with irregular shape or abnormal reflection characteristics have large brightness fluctuations and a standard deviation significantly higher than the average level.

[0038] In the final impurity type determination process, the system makes a judgment based on the analysis results of the differential image and the time series image: when the intensity of a certain area in the normalized differential image exceeds the dynamic judgment threshold, it is determined that there are compositional difference impurities in the area; and when the area does not deviate significantly in the differential image, but shows a high time standard deviation in its corresponding total luminous flux time series, it is determined to be morphological difference impurities. This dual-path determination mechanism fully covers the needs of identifying multiple types of impurities caused by material differences and structural abnormalities, enhancing the comprehensiveness and accuracy of the detection system.

[0039] Example 2: In a typical application scenario, the method is applicable to the recovery and recycling of metal powder in the additive manufacturing process. Under this working condition, the metal powder is in a continuous flow state and may be mixed with foreign particles of complex origin, including residual oxides and inclusions with different compositions from the substrate, and metal chips of the same material but with a broken shape. Therefore, the detection system must not only have high sensitivity, but also be adapted to the real-time operating environment and have robust response capabilities to differences in powder particle size, morphology, and composition. The detection process of this embodiment includes the following five main links: polarized illumination construction, image acquisition and decoding, normalized differential image generation, dynamic threshold analysis and judgment, time-series-based dynamic feature extraction, and final recognition and classification. The key contents of each link are explained in turn below.

[0040] In terms of lighting construction, this system uses a standard industrial LED surface light source and an industrial-grade circular polarizer at the front end. The circular polarizer can convert linearly polarized light into circularly polarized light with phase delay characteristics, has stable polarization direction consistency and optical performance, and is suitable for the lighting needs of dynamic powder channels. Through this setting, the reflection characteristics of metal powder in different microstructural directions can be enhanced, providing a recognizable basic signal for polarization component extraction; the image acquisition part uses an industrial black and white camera, and a polarization splitter prism or a micro-polarization array filter is configured at the front end of its lens. The above components enable the camera to synchronously acquire images of two orthogonal linear polarization directions in different spatial areas of the image sensor during a single exposure process. Specifically, the polarization splitter prism relies on the principle of birefringence to separate the optical path, and the micro-polarization array filter realizes polarization encoding by arranging directional filter units on the surface of the sensor. To ensure the balance of the intensity distribution of the two polarization images in subsequent calculations, the selected splitter device should have a nearly symmetrical splitting ratio characteristic, so that the obtained image is within the dynamic response range of the sensor and minimizes signal bias; in the image processing stage, the system introduces normalized difference image generation Mechanism, each frame of the captured image is divided into two orthogonal polarization component images after decoding, and intensity normalization is performed on each of them. Subsequently, the absolute value of the grayscale difference is calculated for each pixel of the two images, and the sum of the two image intensities plus a preset non-zero small constant is used as the denominator for normalization. The constant is set according to the quantization accuracy of the image signal to avoid calculation errors caused by division by zero. The normalized difference image can effectively suppress the influence of uneven illumination, and significantly enhance the light intensity change of the impurity area caused by the difference in polarization response, providing a signal basis for subsequent judgment; at the threshold During the value generation and recognition stage, the system constructs a pixel intensity histogram based on the normalized differential image, and calculates the mean and standard deviation of the main distribution. The main distribution is considered to correspond to the intensity background of qualified powder particles in the image. Combining engineering experience and image characteristics, the system dynamically sets the judgment threshold as the mean of the main distribution plus a preset coefficient multiplied by the standard deviation. The coefficient usually ranges from five to six. When the differential intensity of a region in the image is higher than the dynamic threshold, the region is preliminarily marked as a potential impurity. This judgment strategy does not rely on a pre-trained model and has good environmental adaptability and engineering maintainability.

[0041] In order to identify particles with unclear signals but abnormal morphology in the differential image, the system introduces a time-series image sequence analysis path. This process reuses the weighted superposition results of the two polarization component images mentioned above to construct a total luminous flux image and form a time-series image sequence in continuous frames. Subsequently, based on the displacement information of the powder particles, the position of specific particles is tracked by optical flow or regional correlation matching, and a dynamic observation window is established on its path. Within this observation window, the system counts the brightness values ​​of each pixel in the image sequence over time and calculates its time standard deviation. Under normal circumstances, regular particles have small brightness fluctuations due to stable reflection, while debris with broken morphology and irregular surface exhibits greater brightness fluctuations due to factors such as posture changes, so its standard deviation is significantly increased, which can be used as a basis for identifying morphological abnormalities. The final judgment process integrates the results of differential image analysis and time series fluctuation analysis. If the intensity of a certain area in the differential image exceeds the dynamic threshold, it is judged as a composition difference impurity; if a certain area does not exceed the differential threshold, but presents a time standard deviation higher than the background benchmark in the corresponding time series image, it is judged as a morphological difference impurity. This dual-channel recognition mechanism enables the system to cover multiple types of impurity recognition scenarios caused by material differences or morphological anomalies, improving the overall detection coverage and judgment accuracy; the system is deployed on an embedded platform with image acquisition and processing capabilities. Common selections include general-purpose processing modules with image interfaces and medium computing performance. Under this platform, image acquisition, differential processing, dynamic analysis and judgment processes can be fully processed locally without relying on external servers or manual intervention, which is suitable for long-term stable operation on industrial production lines. During implementation, the system does not rely on edge detection, deep learning models or template matching programs, and has engineering advantages such as simple structure, low deployment cost and easy maintenance.

[0042] Example 3: This example aims to explain in depth the specific implementation approach and performance verification of a method for detecting impurities in metal powder based on image recognition. This method aims to address the limitations of traditional technologies in the detection of microscopic impurities through the synergistic effect of polarized light response differences and timing analysis, and to improve the robustness and long-term stability of the system in dynamic industrial environments. In high-tech fields such as additive manufacturing, the purity of metal powder is a key factor in determining the quality of the final product. In actual production, even after preliminary screening, the powder may still be mixed with trace impurities during subsequent processing, including particles with different composition or abnormal morphology from the base powder. Existing detection methods often have difficulty achieving stable and accurate detection when faced with microscopic impurities with extremely low optical contrast or under conditions of image blur caused by high-speed flow, and are prone to missed detections. In addition, factors such as ambient lighting and powder batch differences may cause image signal drift, resulting in traditional detection methods based on fixed thresholds requiring frequent manual calibration, affecting the efficiency and stability of automated production lines. To this end, this technical solution proposes a detection method, the core of which is to utilize the essential differences in the polarization light response between metal powder and impurities to identify compositional impurities, and to capture the dynamic characteristics of morphologically abnormal impurities through time-series analysis of the total light flux image sequence. At the same time, adaptive threshold adjustment is achieved in combination with statistical analysis, aiming to improve the system's adaptability to environmental changes and long-term operational stability.

[0043] This experiment built a verification platform in a simulated industrial powder fluid detection environment. The configuration considerations of the main components include: a polarized illumination unit, which uses an industrial-grade LED surface light source with power adjustment function, and a customized industrial standard circular polarizer is installed at the front end to form a stable circularly polarized illumination field in the detection area and provide uniform illumination; an image acquisition unit, which uses an industrial black and white camera with a resolution and frame rate setting designed to capture high-speed flowing powder images. A micro-polarization array filter is configured in front of the camera lens to enable the camera to synchronously acquire at least two orthogonal linear polarization component images in a single exposure to ensure the real-time and synchronization of image data acquisition; a powder conveying and flow field control unit, which consists of a high-precision vibrating feeder and a slit The system is composed of a flow guide trough, which is used to achieve stable and uniform flow of metal powder. The flow rate is precisely adjusted by a flow controller to simulate the dynamic working conditions that may occur in an actual production line; and a data processing and analysis unit. The core processing module is based on an embedded system development board. The platform has image processing acceleration capabilities and can support high-speed transmission of real-time image data and parallel computing of complex algorithms. In this experiment, 316L stainless steel spherical powder commonly used in additive manufacturing is selected as the matrix material, and alumina micropowder is used as the composition-differential impurity, and 316L stainless steel chips generated by wear are used as the morphology-differential impurities. Both are uniformly doped in the matrix metal powder at trace mass concentrations to truly reproduce the situation of impurity mixing in actual production.

[0044] The experiment was conducted according to the following steps: First, under the condition of pure metal powder flowing without impurities, the polarization illumination unit and image acquisition unit were carefully optically calibrated to ensure that the brightness of the image acquired was moderate and without saturation. The background signal intensity distribution of the pure powder in the polarization difference image was recorded as a benchmark for subsequent adaptive threshold adjustment. Then, metal powder doped with simulated impurities was injected into the detection area and transported. The image acquisition unit synchronously acquired the orthogonal linear polarization component images of the light reflected by the metal powder. The data processing unit performed model-free differential information extraction in real time, that is, each frame of the image was divided into the first linear polarization component image and the second linear polarization component image, and then the normalized differential image intensity was calculated. Its calculation method follows the preset logic, which is the ratio of the absolute value of the intensity difference between the two linear polarization component images to the sum of their intensities (plus a non-zero positive number). The setting of this non-zero positive number is intended to avoid division by zero and ensure effective distinction of weak signals. Its value can be reasonably set with reference to the minimum grayscale response unit of the image sensor. The system obtains the pixel intensity histogram of the normalized difference image in real time, and performs statistical analysis based on this histogram to determine the mean and standard deviation representing the main distribution of qualified metal powder. According to the determined mean and standard deviation, the system dynamically generates a judgment threshold. The method for generating the judgment threshold follows the principle of the mean plus a preset coefficient multiplied by the standard deviation, wherein the selection of the preset coefficient is, Based on comprehensive exploratory tests and statistical analysis of the fluctuation range of optical properties of different powder batches, and combined with the prudent determination of the balance requirement between false alarm rate and missed alarm rate in engineering practice, the area with pixel intensity higher than the judgment threshold in the normalized difference image can be preliminarily identified and marked as a potential area of ​​composition difference impurities; in parallel with the above steps, the system obtains a total luminous flux image sequence composed of the superposition of the first linear polarization component image and the second linear polarization component image. For the potential impurity area that has been preliminarily identified, the system uses a micro-trajectory tracking algorithm based on optical flow method or regional correlation matching to track the motion of powder particles in the area between consecutive frames, so as to establish a dynamic view that follows the movement of the particles. Measurement window: Within the dynamic observation window, the system continuously monitors the temporal variation of pixel brightness values ​​and calculates the temporal standard deviation of these values ​​in a continuous image sequence. This temporal standard deviation reflects the degree of fluctuation in the intensity of reflected light from particles during flow. At the same time, a morphology determination threshold is set. This threshold can be reasonably set based on statistical analysis of light intensity fluctuation characteristics of a large number of powder particles with different morphologies in high-speed flow and engineering experience, aiming to effectively distinguish between particles with stable morphology and particles with abnormal morphology. If the intensity of a certain area in the normalized difference image does not reach the composition determination threshold, but the corresponding total light flux time series analysis shows a temporal standard deviation higher than the morphology determination threshold, then the area is determined to be morphologically different debris.The system records the number of impurities detected, false positives, and missed positives for each batch of tests, and verifies the test results through sampling combined with high-resolution microscopy. The tests are repeated under different ambient lighting and powder flow rate conditions to evaluate the robustness of the system.

[0045] This experiment systematically collected detection data under different working conditions, and the results showed that: in terms of the identification of impurities with composition differences, when detecting aluminum oxide impurities, a high detection rate can be obtained through normalized differential image detection. Qualified metal powder has a small intensity difference in the orthogonal polarization component image due to its optical reflection characteristics, forming a low-intensity background in the normalized differential image. Aluminum oxide particles, as non-metallic impurities, have obvious optical anisotropy in their response to polarized light, resulting in a signal higher than the background in the normalized differential image. This phenomenon verifies the effectiveness of the polarization differential mechanism in improving the detection sensitivity of impurities with microscopic composition differences. In terms of the identification of impurities with morphological differences, for special-shaped steel chips, simple normalized differential detection is not effective, but through the time series analysis of the total light flux image, the detection rate of this type of impurities is significantly improved. When steel chips with regular shapes flow at high speed, their reflected light intensity fluctuates with changes in posture, resulting in a significantly higher time standard deviation of pixel brightness within the dynamic observation window than that of spherical powders. This indicates that time series analysis can effectively capture the changes in dynamic optical properties caused by morphological abnormalities, providing an independent identification basis for morphologically different impurities. The advantage of two-dimensional combined detection is that when compositionally different and morphologically different impurities coexist, the system achieves a high overall detection rate by combining detection to determine that any dimension reaches the threshold and is an impurity. Both the omission rate and the false alarm rate are controlled at a low level, verifying the complementarity of the composition and morphology dual-dimensional detection capabilities proposed in this method, which helps to fully cover complex impurity types. In addition, this experiment verified the adaptive threshold adjustment mechanism under conditions simulating light source brightness fluctuations and powder batch differences in actual industrial environments.

[0046] Experimental results demonstrate that even when the background mean and standard deviation fluctuate with external conditions, the system can track these changes in real time and adjust the dynamic judgment threshold accordingly. This threshold consistently maintains an appropriate distance from the main distribution of qualified powder under the current operating conditions, effectively suppressing false alarms caused by background noise while maintaining sensitivity to impurity signals. This demonstrates the system's adaptability to environmental perturbations and long-term operational stability. To evaluate the real-time processing capabilities of this method on an embedded platform, we tested the average processing latency of the selected embedded processor platform at different image frame rates. The test results show that the average total latency for completing the entire processing of a single frame image on this embedded platform at a typical frame rate of an industrial camera is acceptable for industrial real-time detection applications. Although the timing analysis is computationally intensive, its processing time is kept to the millisecond level and is performed only on the initially identified potential impurity areas, with minimal impact on overall real-time performance. This demonstrates that the selected embedded platform and its optimized algorithm deployment can effectively support high-frame-rate real-time detection requirements and achieve engineering feasibility. The model-free nature of this method reduces excessive reliance on computing resources, helps simplify the system structure, and improves the ease of engineering deployment.

[0047] Example 4: This example combines Figures 1 to 3 , a method for detecting impurities in metal powder based on image recognition is described. Figure 1 As shown, first, in the optical imaging and acquisition system, the industrial LED surface light source forms circularly polarized light through the circular polarizer set at the front end, and irradiates the flowing metal powder in a flowing state. The polarized light reflected by the metal powder is received by the polarization splitting and imaging unit and decomposed into the first and second linear polarization component images. Then, the industrial black and white camera sensor synchronously collects the original polarization image data and enters the core processing and judgment engine stage. The image data is first input into the system processing module through the image data input interface, and then the two-dimensional information analysis module performs image difference calculation and total luminous flux image sequence generation, and inputs the normalized difference image result into the dynamic threshold decision module, which generates a dynamic judgment threshold based on the real-time image statistical features. The system then compares the difference result with the dynamic threshold, and combines the total luminous flux image sequence data to identify the suspected impurity area in the image by the impurity collaborative classification module, and finally outputs it to the detection result output interface.

[0048] like Figure 2As shown, the polarized illumination and acquisition unit consists of an industrial LED surface light source, a circular polarizer, and an industrial black-and-white camera. The industrial LED surface light source is used to provide detection illumination. Its light passes through the circular polarizer to form circularly polarized light with specific phase characteristics, which is then irradiated on the metal powder to collect incident light, that is, to illuminate the dynamically flowing metal powder and collect its reflected light signal. After polarization processing, the reflected light is captured by the industrial black-and-white camera. The collected data is output as image data via a USB or network transmission interface. The image data is input into the embedded processing unit. In this unit, for example, a Raspberry Pi or ESP32-S3 can be used as the main control computing platform responsible for performing image processing tasks. The core image processing algorithm consists of a core processing and judgment algorithm, which includes three parts: an image difference module, a time series analysis module, and an adaptive threshold module. The image difference module is used to calculate the normalized difference image to identify compositionally different impurities. The time series analysis module is used to dynamically analyze the total light flux image sequence to extract morphological change characteristics. The adaptive threshold module dynamically generates recognition and judgment thresholds based on image statistical results to ensure the robustness and detection accuracy of the system in changing environments.

[0049] like Figure 3 As shown in the figure, in the initialization phase, the image processing module first passes part of the image feature data transmitted in the initial phase to the dynamic threshold module. The module generates an initial judgment threshold by analyzing the histogram and calculating the initial mean and standard deviation, returns the initial judgment threshold to the image processing module, and applies the initial threshold to start detection; in the routine operation and monitoring phase, the system continuously monitors the stability of the environment and image features. The system monitoring module periodically detects and processes the image feature recognition results, and uploads the current image feature distribution statistical characteristics for system monitoring and evaluation. When the system monitoring module detects fluctuations in system characteristics (such as aging of the light source or uncorrected environmental disturbances), it will trigger a re-update control instruction, which requests the dynamic threshold module to re-output the histogram information of the current statistical image; then, the dynamic threshold module will re-analyze the histogram and calculate the new mean and standard deviation, and return the updated judgment threshold to the image processing module. The image processing module will apply the updated judgment threshold to continue detection; if no difference is detected, the system confirms that the current threshold is valid.

[0050] like Figure 4As shown in the figure, the original image shows the imaging picture of the metal powder in the flowing state, which serves as the input of the differential analysis; the polarization difference image shows the image data generated by calculating the normalized polarization component difference, highlighting the abnormal particles with optical anisotropy to polarized light; in the detection result annotation diagram, the identified impurity targets are classified and labeled by image visualization, where C: composition difference indicates that the detected composition difference impurities are marked with red circles, and M: morphological difference indicates that the detected morphological difference impurities are marked with yellow boxes. The statistical information at the bottom of the figure clearly indicates that the number of impurities detected is 5, of which 3 are composition difference type and 2 are morphological difference type, indicating that the technical solution of the present invention effectively realizes impurity identification and classification through a two-dimensional collaborative mechanism. In the detection parameter section, the dynamic threshold calculation formula used in the current detection is listed: , which clearly reflects the adaptive calculation logic of the dynamic judgment threshold; the image acquisition frame rate is 120fps, which ensures the full acquisition of image data under fast flow conditions; the detection area is 10mm×7.5mm, which indicates the size of the detection window; the powder type is 316L stainless steel spherical powder with a particle size range of 15–45μm, which describes the experimental conditions.

[0051] Example 5: In this example, a detection scheme for impurities in metal powder is deployed on a raw material pretreatment production line for the production of high-precision powder metallurgy parts. The production line is equipped with an image acquisition and recognition system. Combined with the synchronization logic of the conveyor belt movement and the image acquisition window, it can realize frame-by-frame imaging and dynamic recognition of batch metal powders. The system integrates a high-frame rate, narrow-bandwidth industrial-grade camera and an imaging cabin with a shading structure. When the metal powder passes through a preset illumination area, the powder microscopic image is acquired at a set frequency. The acquired image is grayscale balanced, noise suppressed, and edge enhanced by the image preprocessing module to meet the requirements of the downstream recognition model for input data quality. It should be pointed out that in actual deployment, the sampling rhythm of the imaging window needs to be dynamically matched with the operating speed of the conveying system. For this reason, the control system is based on According to the real-time feedback of the encoder on the belt running speed, the imaging trigger frequency is automatically adjusted to ensure complete coverage of the same powder batch and non-overlapping between images; the image recognition part adopts a multi-level image feature extraction structure, combined with a multi-scale discrimination mechanism based on the fusion of local contrast and edge contour response to achieve robust detection of non-metallic impurities such as oxide flakes, fibrous particles, siliceous debris, etc. Specifically, the module first extracts the low-frequency components and local texture indicators in each frame of the image, and then constructs its edge intensity map and texture response map at multiple spatial scales. On this basis, the system introduces a class of discriminant functions based on hierarchical responses to quantify the difference index between the suspicious area and the standard metal particle area in the image. This index not only takes into account the local contrast difference, but also integrates the regional edge integrity and the degree of deviation of the grayscale distribution.

[0052] In order to improve the adaptability of the discrimination mechanism in actual scenarios, this embodiment further introduces a set of dynamically adjusted discrimination thresholds, which are adaptively adjusted according to the overall grayscale mean and texture complexity of each batch of metal powder. When the system first processes the image stream, it first constructs a statistical distribution model of the current batch of images, including the grayscale histogram and edge complexity distribution, and then derives the discriminant function adjustment factor applicable to this batch in combination with the preset response mapping relationship, to ensure that the detection algorithm has consistent and stable recognition sensitivity between powder samples from different batches or different supply sources; after detecting a suspected debris area, the system automatically generates a mask with coordinate markings in the original image, and sends the image and detection results of the corresponding area to the quality judgment module through a control signal. The module combines historical image patterns, typical debris morphology and false detection tolerance strategy to further determine whether the current detection result constitutes a real abnormal event. If it is determined to be valid debris, the system will trigger the rejection device action signal to control the subsequent mechanical sorting mechanism to physically sort the powder segment. Isolate to prevent it from entering the next process; it is worth noting that the data interaction between each module in the image processing process is completed through a standardized data interface, and the transmission format includes a uniformly defined intermediate image data structure and an asynchronous annotation instruction protocol to ensure data compatibility and timing consistency between each sub-module. For example, in this embodiment, the data transmission between the image recognition module and the quality judgment module adopts a compressed regional response vector and a reversible mask image structure to achieve fast transmission and low-latency collaborative judgment without losing spatial information. In addition, in order to improve the stability of the entire recognition mechanism under actual working conditions, this embodiment introduces a system state adjustment mechanism based on environmental perception. This mechanism monitors changes in the imaging environment in real time through the deployment of temperature and humidity sensors and conveyor belt vibration sensors, and adjusts the light source intensity, imaging parameters and filtering strategies when necessary. For example, when the environmental humidity increases and the image contrast decreases, the system automatically increases the light source brightness and adjusts the exposure time to ensure that the image clarity remains within the effective recognition range.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting impurities in metal powder based on image recognition, characterized in that: The method comprises the following steps: Step a: constructing a polarized illumination field, wherein an industrial LED surface light source is used, and a circular polarizer is arranged in front of the industrial LED surface light source to irradiate the flowing metal powder with circularly polarized light; Step b: synchronously capturing at least two orthogonal linear polarization component images of light reflected from the metal powder, wherein an industrial black-and-white camera is used, and a polarization beam splitter prism or a micro-polarization array filter is disposed in front of the lens of the industrial black-and-white camera, so that the camera can receive at least two orthogonal linear polarization component images in different areas of its single image sensor in a single exposure; Step c, performing model-free differential information extraction, wherein each frame of image captured by the camera is processed, the image is divided into a first linear polarization component image and a second linear polarization component image, and a normalized differential image is calculated, where the intensity of each pixel of the normalized differential image is the absolute value of the intensity difference of the pixel in the first linear polarization component image and the second linear polarization component image, divided by the sum of the intensity of the first linear polarization component image, the intensity of the second linear polarization component image, and a non-zero positive number; Step d, identifying pixel areas in the normalized difference image whose pixel intensity is higher than a preset determination threshold as impurities in the metal powder; Step e: acquiring a time-series image sequence consisting of a total luminous flux image formed by superimposing a first linear polarization component image and a second linear polarization component image, wherein the total luminous flux image intensity = the first linear polarization component image intensity + the second linear polarization component image intensity; Step f: performing micro-trajectory tracking of moving particles in the time-series image sequence using an optical flow method or correlation matching to form a dynamic observation window that follows the movement of the particles; and Step g: analyzing the temporal variation characteristics of pixel brightness values ​​within the dynamic observation window, wherein the temporal standard deviation of the temporal image sequence within the dynamic observation window is calculated; and identifying impurities with non-spherical morphological characteristics in the metal powder based on the temporal standard deviation; In step g, if an area shows an intensity higher than the judgment threshold on the normalized difference image, it is judged as a composition difference type impurity; if an area shows an intensity lower than the judgment threshold on the normalized difference image, but shows a time standard deviation higher than the preset morphological threshold in its corresponding total luminous flux time series analysis, it is judged as a morphological difference type impurity.

2. The method for detecting foreign matter in metal powder based on image recognition according to claim 1, characterized in that: In step d, the identification includes: obtaining a pixel intensity histogram of the normalized difference image; performing statistical analysis on the pixel intensity histogram to determine the mean value representing the main distribution of qualified metal powder and standard deviation ; and based on the mean and standard deviation , dynamically generate the decision threshold, where the decision threshold , It is a preset coefficient with a value range of 5 to 6.

3. The method for detecting foreign matter in metal powder based on image recognition according to claim 2, characterized in that: The splitting ratio of the polarization beam splitter prism or the micro-polarization array filter enables at least two orthogonal linear polarization component images to obtain a balanced intensity distribution within the dynamic response range of the camera sensor.

4. The method for detecting foreign matter in metal powder based on image recognition according to claim 3, characterized in that: The normalized differential image calculation in step c is completed in real time on the embedded processor.

Images