A method for image recognition of ductile iron based on ai and its computer-readable medium.

Abstract

A method for image recognition of ductile iron based on artificial intelligence and a computer-readable medium thereof are disclosed. The method includes the following steps: capturing an image; training an artificial intelligence model; recognizing the image; and analyzing the image. The present invention further provides a computer-readable medium containing instructions for executing the method of the present invention. In this invention, deep learning training of an artificial intelligence model is used to identify ductile iron particles and improve the efficiency and accuracy of calculating the spheroidization rate.

Description

Technical Field

[0001] This invention relates to an image recognition method using artificial intelligence, and more particularly to the image recognition and analysis of spheroidal graphite cast iron by artificial intelligence after training. Prior Technology

[0002] Ductile cast iron, a type of metallographic structure, possesses high impact resistance and fatigue resistance, and is widely used in industrial fields. Due to the increasing demands for high-value, precision, and functionality in products, metallographic analysis is a primary criterion for judging the quality of components. Currently, manual inspection is one of the main methods for analyzing metallographic images in the industry. This requires skilled technicians to interpret and analyze the metallographic images, necessitating significant investment of time and money in training them. However, manual inspection involves many uncertainties, such as the individual technician's experience, mental state, and judgment standards. Furthermore, manually analyzing a single image can take anywhere from 30 minutes to an hour, and often multiple images are taken for analysis. Excessive analysis time can also affect the technician's interpretation of the metallographic images.

[0003] Automated Optical Inspection (AOI) is also a method for metallographic analysis. AOI uses a rule-based judgment mechanism, relying on pre-written program logic, but is limited by parameter definitions, which narrows the scope of sample analysis and often leads to missed detections. However, due to limitations in program settings and the susceptibility of the analyzed images to factors such as lighting and sample characteristics, even slight differences from the set parameters can result in significant discrepancies between the output image and the actual image analysis data. Therefore, most metallographic image analysis still relies on manual verification.

[0004] Chinese patent number CN108647718B (Patent title: A method for automatic classification and grading of metallographic structures of different materials) discloses an automatic classification and grading method for different metallographic materials. It is based on a convolutional neural network algorithm to establish a database of grain levels for three different materials at different magnifications. After classifying the database, the image data is enhanced, fitted with a specified function, and then transfer learning is used to automatically grade the metallographic images. However, this previous patent lacked practical grading examples and failed to demonstrate the accuracy and efficiency of other technologies in its implementation.

[0005] It is evident that manual verification requires a considerable amount of time to analyze images and relies heavily on the experience of technicians, as well as differing judgment standards among individual technicians. These factors negatively impact the image results of metallographic analysis. Existing metallographic analysis detection systems are limited by difficulties in parameter adjustment and the inability to identify different types of metallographic images. Improving the image recognition efficiency of ductile iron castings in metallographic structures and increasing the accuracy of spheroidization rate within the ductile iron particles in the images are currently among the problems that those skilled in the art wish to solve. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide an image recognition method for identifying ductile iron particles by training an artificial intelligence model through deep learning, thereby calculating the spheroidization rate, so as to improve the accuracy of spheroidization rate recognition and reduce detection time.

[0007] To achieve the aforementioned objective, this invention provides an image recognition method for identifying ductile iron based on artificial intelligence, comprising the following steps: image acquisition; providing a sample to be tested and acquiring the sample through a photographic device to obtain an original image; inputting the original image into an artificial intelligence model; training the artificial intelligence model; the artificial intelligence model receiving the original image and performing a classification process to generate at least one group; the artificial intelligence model generating a training set after training and testing the group division from these groups; extracting the original image from the training set; performing image preprocessing on the original image to obtain an enhanced image; inputting the original image and the enhanced image into the artificial intelligence model to perform recognition and generate a result image. The system uses a threshold to define a region of the resulting image, wherein the image recognition of the artificial intelligence model is trained by deep learning on the input original image to generate an optimized artificial intelligence model; image identification is performed by the optimized artificial intelligence model to identify the resulting image based on the original image and the enhanced image, and the region of the resulting image is defined by the threshold, and at least one spheroidal graphite particle is identified by an instance segmentation algorithm; and image analysis is performed by identifying the spheroidal graphite particles by the instance segmentation algorithm, and obtaining a maximum diameter, an area and a maximum circumscribed circle area of ​​the spheroidal graphite particles, thereby calculating a roundness coefficient and a spheroidization rate, wherein the roundness coefficient is used to distinguish the morphology of the spheroidal graphite particles.

[0008] In some embodiments, the artificial intelligence is built upon the Mask Region-Based Convolutional Neural Networks (MASK R-CNN) algorithm model architecture based on the instance segmentation algorithm.

[0009] In some embodiments, the threshold is used to define the region based on the resulting image identified by the artificial intelligence model.

[0010] In some embodiments, the image preprocessing further includes the following steps: forming a grayscale image by converting the original image to grayscale, wherein forming the grayscale image is to remove the influence of color on the original image and to save data storage space and memory; forming a contour recognition image by performing Gaussian blur processing to reduce noise in the grayscale image and performing softening processing to form a contour recognition image; and forming an enhanced image by performing binarization processing on the contour recognition image to obtain the enhanced image, wherein the enhanced image is to better adjust the pixel grayscale and threshold settings, which helps the artificial intelligence model to identify objects in the image, wherein the objects are the spherical particles.

[0011] In some embodiments, the area of ​​the largest circumscribed circle is calculated from the largest diameter, and the roundness coefficient is calculated using the following formula (1); Formula (1): Based on the roundness coefficient, the model is divided into three different ranges: a first model, a second model, and a third model. The roundness coefficient of the first model is less than 48%, the roundness coefficient of the second model is greater than or equal to 58% and less than 68%, and the roundness coefficient of the third model is greater than or equal to 68%.

[0012] In some embodiments, the maximum diameter of the ductile iron particle is less than 15 µm and is not included in the roundness coefficient calculation.

[0013] In some embodiments, the sphericity is calculated using the following formula (2); Formula (2):

[0014] In some embodiments, the method further includes the following steps: outputting an image, wherein the optimized artificial intelligence model inputs data to a display interface and outputs the data on the display interface, wherein the data includes the original image and the result image, and the spherical particles in the result image are marked with colors to distinguish the different types of spherical particles, and the colors are red, yellow or green.

[0015] In some embodiments, the data is presented in tabular form, which further includes the result image, the spheroidization rate, the number of ink particles representing the first type, the second type, and the third type, and the total number of ink particles.

[0016] The present invention provides a computer-readable medium, which is used in a computing device or computer, and stores instructions to execute the image recognition method based on artificial intelligence for metallographic recognition as described above.

[0017] This invention discloses an image recognition method for ductile iron based on artificial intelligence and its computer-readable medium. By using an artificial intelligence model for deep learning training, the method identifies the size of ductile iron particles in the image and calculates the spheroidization rate, thereby reducing the time cost of relying on manual inspection and metallographic analysis and improving the accuracy of spheroidization rate identification. Simple Explanation of the Diagram

[0018] Figure 1 is a flowchart of the steps of an image recognition method for ductile iron based on artificial intelligence according to the present invention. Figure 2 is a partial flowchart of the image recognition method for identifying ductile iron based on artificial intelligence according to the present invention. Figure 3 is a partial flowchart of the image recognition method for ductile iron based on artificial intelligence according to the present invention. Figure 4 is a partial flowchart of the image recognition method for ductile iron based on artificial intelligence according to the present invention. Figure 5 shows the regional hotspot map of ductile iron particles within the region identified by the optimized artificial intelligence model of this invention. Figure 6 shows the maximum diameter, area, and maximum circumscribed circle area of ​​the ductile iron particles of the present invention. Figure 7 shows how color is used to distinguish the morphology of ductile iron particles in the image recognition of this invention. Implementation

[0019] The following specific embodiments illustrate the implementation of the present invention, providing a more detailed description. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different embodiments, and various details in this specification can be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0020] The invention described in the embodiments disclosed in this specification is an image recognition method for identifying ductile iron particles by training artificial intelligence, thereby calculating the spheroidization rate to improve the accuracy of spheroidization rate recognition and reduce detection time.

[0021] The present invention discloses an image recognition method for ductile iron based on artificial intelligence, as shown in Figure 1 and partial flowcharts of Figures 2 to 4. Figure 1 is a flowchart of the steps of the image recognition method for ductile iron based on artificial intelligence of the present invention, which mainly includes the following steps:

[0022] To further illustrate how the test samples in this invention are obtained, the following description, along with accompanying figures, will provide an explanation.

[0023] Prepare the sample. After hot embedding the sample to be tested, grind and polish it to obtain the ductile iron sample.

[0024] As shown in Figure 2, the sample preparation in S10 further includes the following steps: S110 hot embedding involves placing the ductile iron sample in a hot embedding machine and applying heat and pressure to embed the sample, ensuring the edge integrity of the ductile iron sample. After S120 grinding and hot embedding, the ductile iron sample was removed and one side of the ductile iron sample was ground with wet sandpaper. S130 polishing involves applying alumina powder to the grinding surface of the ductile iron sample for polishing; and S140 sampling involves rinsing the ductile iron sample with water and alcohol, followed by drying to obtain the sample to be tested.

[0025] The sample preparation step follows a common sampling procedure for ductile iron, which is understandable and operable by those skilled in the art to which this invention pertains.

[0026] S10 Image Acquisition: The sample to be tested is obtained during the sample preparation step, and an image of the sample is captured using a photographic device. This original image is then input into the artificial intelligence model. Specifically, the photographic device is preferably an optical microscope. The sample to be tested is placed on the stage of the optical microscope, which preferably observes the sample at 100x magnification. The image of the sample is automatically captured, and the original image is transmitted to the artificial intelligence model. In some embodiments, the original image is annotated.

[0027] As shown in Figure 3, the process for training the artificial intelligence model in S20 is as follows: The S210 artificial intelligence model receives raw images; The S230 artificial intelligence model performs classification processing to generate at least one group. After training and testing the group, a training set is generated in the artificial intelligence model. The original image is extracted from the training set, and the original image is preprocessed to obtain the enhanced image. The original image and the enhanced image are input into the artificial intelligence model to perform recognition and generate the result image. The area range of the S250 threshold segmentation result image; and The S270 artificial intelligence model's image recognition system is trained using deep learning by inputting raw images, thereby generating an optimized artificial intelligence model.

[0028] As shown in Figure 4, image preprocessing in training the S230 artificial intelligence model includes the following steps: S235 generates a grayscale image by converting the original image to grayscale. Converting a color image to grayscale reduces the influence of color on the original image, allowing AI models to focus more on brightness variations, saving data storage space and memory, and accelerating image processing while reducing computational costs. The S240 generates a contour recognition image. Gaussian blurring is applied to reduce noise in the grayscale image and soften it, thus forming the contour recognition image. Gaussian blurring is used to remove noise from the grayscale image, making it easier to obtain accurate image contours, and to adjust the grayscale and threshold values ​​required for calculation in the artificial intelligence model, aiding in the detection of objects or specific features in the image; and S245 generates an enhanced image, performs binarization processing on the contour recognition image, and obtains the enhanced image. The binarization processing is to better adjust the pixel grayscale and threshold settings, which helps the artificial intelligence model to identify objects in the image, where the objects are spherical particles.

[0029] Image preprocessing primarily aims to improve the quality of the original image by extracting and processing the necessary information, thereby enhancing the image for image recognition in artificial intelligence models.

[0030] Step S20 explains that the artificial intelligence is built upon the Mask Region-Based Convolutional Neural Network (MASK R-CNN) algorithm model architecture based on instance segmentation. It performs sphericity classification processing to generate at least one sphericity level. This "at least one" refers to the fact that the original image, due to differences in brightness and shooting angle, forms images with different appearances. To enable the artificial intelligence model to better recognize and classify images, different image groups are generated. The original images are extracted from different groups and preprocessed to transform them into enhanced images, allowing the artificial intelligence model to better identify objects in the images. The original and enhanced images are input into the artificial intelligence model to perform image recognition and generate a result image. A threshold is used to divide the region range of the result image; different region ranges are divided based on the different result images identified. This process involves inputting a large number of original images into the artificial intelligence model for deep learning training, thereby generating an optimized artificial intelligence model.

[0031] Specifically, the Mask Region-Based Convolutional Neural Networks (MASK R-CNN) algorithm model is mainly derived from Faster R-CNN. It adds a Mask branch to handle instance segmentation and predict the category of each pixel. The backbone uses a ResNet-101 network, which helps with the detection of small objects. It uses the RoI Align method to improve the accuracy loss caused by the rounding of ROI Pooling. It is combined with FPN to enhance the feature extraction capability, and has good detection and instance segmentation performance.

[0032] S30 identifies the image by using an optimized artificial intelligence model to generate a result image based on the original image and the enhanced image. A threshold is used to divide the result image into regions containing at least one spheroidal particle. An instance segmentation algorithm is then used to identify each spheroidal particle. The image identification method is as described in step S30: after dividing the region using a threshold, the instance segmentation algorithm identifies individual spheroidal particles for subsequent image analysis.

[0033] Figure 5 shows a regional hotspot map of ductile iron particles identified within a region using an optimized artificial intelligence model.

[0034] The S40 analyzes the image as shown in Figure 6. It uses the optimized artificial intelligence model to detect ductile iron particles and obtains the maximum diameter D, area A, and maximum circumscribed circle area CA of the ductile iron particles. It then calculates the roundness coefficient and spheroidization rate, where the roundness coefficient is used to distinguish the morphology of the ductile iron particles.

[0035] The area CA of the largest circumscribed circle is calculated from the largest diameter D, and the roundness coefficient is obtained by the following formula (1); Formula (1): By defining roundness coefficient standards for different types, the types are divided into three different ranges: a first type, a second type, and a third type. The roundness coefficient of the first type is less than 48%, the roundness coefficient of the second type is greater than or equal to 58% and less than 68%, and the roundness coefficient of the third type is greater than or equal to 68%.

[0036] The maximum diameter D of the ductile iron particles is less than 15 µm and is not included in the roundness coefficient calculation to avoid the presence of intermediaries affecting the roundness coefficient in the image analysis by the artificial intelligence model. The intermediaries refer to non-metallic components and are not the target objects that the artificial intelligence model of this invention aims to identify.

[0037] The sphericity rate is calculated by calculating the total number of particles of different types, and the sphericity rate is obtained by the following formula (2); Formula (2): The roundness coefficient defines the range of values ​​for different types: the roundness coefficient for the first type is less than 48%, the roundness coefficient for the second type is greater than or equal to 58% and less than 68%, and the roundness coefficient for the third type is greater than or equal to 68%. The sphericity rate is calculated by statistically analyzing the total number of particles of different types.

[0038] Since ductile iron particles are not perfectly spherical, the roundness coefficient is used to determine the sphericity of the particles. The spheroidization rate represents the distribution ratio of ductile iron particles, and its value will affect the wear resistance, toughness, tensile strength and ductility of ductile iron.

[0039] As mentioned above, another embodiment further includes the following steps: outputting an image, the optimized artificial intelligence model inputting data to a display interface, and outputting the data on the display interface. The data includes an original image and a result image, with the spheroidal particles in the result image marked by color to distinguish their morphology. The colors are red (310), yellow (320), or green (330), with different morphology ranges defined based on the roundness coefficient value. As explained in the previous embodiment, and without repetition, the first morphology is red, the second is yellow, and the third is green. Specifically, the artificial intelligence model inputs data to a display interface, preferably a screen, and displays the original image and result image on the screen for comparison and review, as well as the spheroidization rate calculation results.

[0040] As shown in Figure 7, the different types of spherical particles detected on the screen are distinguished by color. The range values ​​of different types are defined according to the roundness coefficient. The first type is red (310), the second type is yellow (320), and the third type is green (330) to examine the range of spherical particles in the image according to their type classification. The color distinction helps to better understand the range distribution of different types of spherical particles in the image.

[0041] The data is presented in tabular form and includes result images, spheroidization rate, number of spheroid particles (for the first, second, and third types), and total number of spheroid particles.

[0042] Confusion Matrix - A Metric for Evaluating Artificial Intelligence Models A confusion matrix is ​​a table representing the prediction results of a learned model, particularly suitable for classification problems. It is typically an N x N matrix, where N is the number of classes. Commonly seen in binary classification, a confusion matrix is ​​a 2 x 2 matrix, as shown in Table 1. Actual Values True Negative (False) predict (Predicted Values) True TP prediction was correct. FP prediction error Negative (False) FN prediction error TN prediction is correct Table 1. Concept of Confusion Matrix Where True Positive (TP) represents a true positive, indicating that the model correctly classifies a positive sample as a positive sample; False Positive (FP) represents a false positive, indicating that the model incorrectly classifies a negative sample as a positive sample; True Negative (TN) represents a true negative, indicating that the model correctly classifies a negative sample as a negative sample; and False Negative (FN) represents a false negative, indicating that the model incorrectly classifies a positive sample as a negative sample.

[0043] A total of 100 images of ductile iron were used to test the artificial intelligence model disclosed in this invention for identifying ductile iron particles. The results are shown in Table 2: (n=100) Actual Values True Negative (False) predict (Predicted Values) True 54 0 Negative (False) 6 40 Table 2. Evaluation of the Artificial Intelligence Model of the Invention As shown in Table 2, the accuracy of the artificial intelligence model reached 94%, the precision rate reached 100%, and the F1 score was 94.7%.

[0044] To verify the feasibility of the artificial intelligence model disclosed in this invention, the spheroidization rate results calculated manually and by the artificial intelligence model were compared. Twenty metallographic samples were randomly selected, with five images for each sample, totaling 100 images. The average spheroidization rate (n=100) was calculated from these 100 images. The average spheroidization rate calculated manually was 74%, while the average spheroidization rate calculated by the artificial intelligence model disclosed in this invention was 76%, resulting in an error rate of 2%. Manually detecting and analyzing a single image takes anywhere from thirty minutes to an hour, while the artificial intelligence model disclosed in this invention can analyze an image and calculate the spheroidization rate within thirty seconds. This experiment demonstrates the feasibility of this invention and its advantage of significantly reducing time costs.

[0045] The present invention also provides a computer-readable medium for an image recognition analysis method based on artificial intelligence metallographic identification, which is applied in a computing device or computer. Preferably, the computer stores instructions and executes the computer-readable medium to perform the image recognition method based on artificial intelligence metallographic identification as described above. Furthermore, each computer within the scope of the patent application of the present invention can implement it using any programming language.

[0046] The present invention discloses an image recognition method for ductile iron based on artificial intelligence and its computer-readable medium. By using deep learning training of an artificial intelligence model, the method identifies the size of ductile iron particles in the image and calculates the spheroidization rate, thereby reducing the time cost of relying on manual inspection and metallographic analysis and improving the accuracy of spheroidization rate identification.

[0047] In summary, this invention addresses the shortcomings of prior art and achieves its main objective. The descriptions of preferred embodiments or examples of the technical means employed to solve the problem are merely intended to limit the scope of this invention. All variations and modifications consistent with the wording of this patent application or made within the scope of this patent are covered by the patent scope of this invention.

[0048] S10~S270: Steps 310: Red 320: Yellow 330: Green D: Maximum diameter A: Area CA: Area of ​​the largest circumcircle

[0049] none

Claims

1. A method for image recognition of ductile iron based on artificial intelligence, comprising the following steps: preparing a sample by hot embedding a ductile iron sample, followed by grinding and polishing to obtain a sample to be tested; capturing an image by capturing the sample to be tested through a photographic device to obtain an original image, and inputting the original image into an artificial intelligence model; training the artificial intelligence model by receiving the original image and performing a classification process to generate at least one group, generating a training set by training and testing the group division of the artificial intelligence model from the groups, extracting the original image from the training set, performing an image preprocessing on the original image to obtain an enhanced image, inputting the original image and the enhanced image into the artificial intelligence model to perform recognition to generate a result image, and a threshold to divide a region of the result image, wherein the image recognition of the artificial intelligence model is trained by deep learning through inputting the original image to generate an optimized artificial intelligence model; The image identification process involves using an optimized AI model to generate a result image based on the original image and the enhanced image. A threshold is used to define the region within the result image, and an instance segmentation algorithm is used to identify at least one spheroidal particle. The optimized AI model can classify the degree of spheroidization based on the result image to generate at least one degree of spheroidization. The image analysis process involves identifying the spheroidal particles using the instance segmentation algorithm and obtaining a maximum diameter, an area, and a maximum circumscribed circle area for each spheroidal particle. A roundness coefficient and a spheroidization rate are then calculated, with the roundness coefficient used to distinguish the morphology of the spheroidal particles.

2. A method for image recognition of ductile iron based on artificial intelligence as described in claim 1, wherein the artificial intelligence is established based on the Mask Region-Based Convolutional Neural Network (MASK R-CNN) algorithm model architecture of instance segmentation algorithm.

3. A method for image recognition of ductile iron based on artificial intelligence as described in claim 1, wherein the threshold is used to delineate the region range based on the resulting image recognized by the artificial intelligence model.

4. The image recognition method for ductile iron based on artificial intelligence as described in claim 1, wherein the image preprocessing further includes the following steps: forming a grayscale image by converting the original image to grayscale to form a grayscale image; forming a contour recognition image by performing Gaussian blur processing to reduce noise in the grayscale image and performing softening processing to form a contour recognition image; and forming an enhanced image by performing binarization processing on the contour recognition image to obtain the enhanced image.

5. A method for image recognition of ductile iron based on artificial intelligence as described in claim 1, wherein the area of ​​the largest circumscribed circle is calculated from the largest diameter, and the roundness coefficient is calculated using the following formula (1); Formula (1): Wherein, Based on the roundness coefficient, the model is divided into three different ranges: a first model, a second model, and a third model. The roundness coefficient of the first model is less than 48%, the roundness coefficient of the second model is greater than or equal to 58% and less than 68%, and the roundness coefficient of the third model is greater than or equal to 68%.

6. A method for image recognition of ductile iron based on artificial intelligence as described in claim 5, wherein the maximum diameter of the ductile iron particles is less than 15 µm and is not included in the roundness coefficient calculation.

7. A method for image recognition of ductile iron based on artificial intelligence as described in claim 1, wherein the spheroidization rate is calculated using the following formula (2); Formula (2): .

8. A method for image recognition of ductile iron based on artificial intelligence as described in claim 1, further comprising the following steps: outputting an image, wherein the optimized artificial intelligence model inputs data to a display interface, outputs the data on the display interface, wherein the data includes the original image and the result image, and the ductile iron particles in the result image are marked with a color to distinguish the different types of the ductile iron particles, the color being red, yellow or green.

9. A method for image recognition of ductile iron based on artificial intelligence as described in claim 8, wherein the data is exported in tabular form, and the data further includes the resulting image, the spheroidization rate, the morphology, and the total number of ductile iron particles.

10. A computer-readable medium, applied in a computing device or computer, storing instructions for performing an image recognition method for identifying ductile iron based on artificial intelligence, as described in any one of claims 1 to 9.

Images