A method for quality inspection of PDC composite sheet molds

Abstract

This application relates to the field of image processing technology and discloses a method for quality inspection of PDC composite sheet molds. The method includes: scanning the PDC composite sheet mold with a multi-frequency ultrasonic probe to obtain multi-channel ultrasonic images, and using a deep convolutional neural network for noise reduction and contrast enhancement. An improved MesoNet algorithm is used to extract features, and a weighted fusion strategy is employed to generate enhanced images. Finally, quality level classification is performed through image segmentation and defect area calculation. This application solves the technical problems of difficult multi-frequency ultrasonic data fusion, inaccurate defect feature extraction, insufficient image segmentation accuracy, and low standardization of quality grading in the quality inspection of PDC composite sheet molds. Through multi-frequency ultrasonic probe collaborative scanning, improved MesoNet algorithm feature extraction, U-Net image segmentation, and intelligent quality classification technology, the accuracy and automation level of defect detection in PDC composite sheet molds are improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method for quality inspection of PDC composite sheet molds. Background Technology

[0002] Polycrystalline diamond composite (PDC) is a special superhard material synthesized in a single process from synthetic diamond micropowder and cemented carbide under high temperature and pressure. It is widely used in cutting tools in fields such as oil drilling and geological exploration. Existing methods for inspecting the quality of PDC composite molds primarily employ ultrasonic testing technology. This involves using an ultrasonic scanning microscope to detect internal defects in the mold, utilizing the reflection characteristics of ultrasound at the interface between the diamond and cemented carbide layers to identify defects such as cracks, voids, and incomplete welds. Traditional inspection methods typically use a single-frequency ultrasonic probe combined with manual judgment, evaluating mold quality based on changes in ultrasonic intensity and waveform diagrams.

[0003] However, existing technologies have significant shortcomings: low detection efficiency, requiring shutdown and mold disassembly for offline inspection, affecting production continuity; limited detection accuracy, single-frequency ultrasonic probes cannot fully cover defect types of different sizes, and manual judgment has subjective and consistency issues; poor real-time performance, making it impossible to detect quality problems in a timely manner during production, often only discovering mold defects after product abnormalities occur; and insufficient data processing capabilities, lacking intelligent image processing and defect recognition algorithms, making it difficult to achieve automated quality assessment.

[0004] Based on an in-depth analysis of the current technological status, the core technical challenges facing the quality inspection of existing PDC composite molds exhibit a progressive characteristic: First, how to obtain more comprehensive and accurate defect information, as the limitations of single-frequency ultrasonic testing prevent the simultaneous and effective detection of deep large defects, mid-layer interface defects, and shallow small defects; second, how to intelligently process complex ultrasonic image data, as traditional manual judgment methods cannot meet the demands of multi-channel and multi-layer image information fusion; this leads to the problem of how to achieve accurate defect localization and quantitative analysis, i.e., how to convert fuzzy ultrasonic responses into precise defect region segmentation and area calculation; and finally, how to establish an objective and reliable quality rating system to transform quantitative defect data into standardized quality level judgments. Summary of the Invention

[0005] This application provides a quality inspection method for PDC composite sheet molds, addressing technical problems such as difficulties in multi-frequency ultrasonic data fusion, inaccurate defect feature extraction, insufficient image segmentation accuracy, and low standardization in quality grading during PDC composite sheet mold quality inspection. By employing multi-frequency ultrasonic probe collaborative scanning, improved MesoNet algorithm feature extraction, U-Net image segmentation, and intelligent quality classification technology, the accuracy and automation level of defect detection in PDC composite sheet molds are enhanced.

[0006] Firstly, this application provides a method for quality inspection of PDC composite sheet molds, the method comprising:

[0007] Step S1: Obtain multi-channel raw ultrasound images by scanning the PDC composite mold with a multi-frequency ultrasound probe;

[0008] Step S2: Input the multi-channel original ultrasound image into a deep convolutional neural network for noise reduction and contrast enhancement to generate a preprocessed ultrasound image, in which the diamond region and the cemented carbide region exhibit different grayscale value distributions.

[0009] Step S3: The preprocessed ultrasound image is feature extracted by improving the MesoNet algorithm. Asymmetric convolution kernels are used to extract interface linear features, and multi-scale convolution kernels are used to detect different types of defect features, and the defect feature vector is output.

[0010] Step S4: Establish a weighted fusion strategy based on the defect feature vector, perform pixel-level fusion of the multi-channel data corresponding to the preprocessed ultrasound image to generate an enhanced fusion image, and obtain a binary defect segmentation mask through an image segmentation algorithm.

[0011] Step S5: Calculate the percentage of defect area to total interface area based on the binary defect segmentation mask, and classify the PDC composite sheet mold quality into four quality levels: high-quality, qualified, substandard, and scrap, according to the defect rate threshold.

[0012] Optionally, step S1 includes:

[0013] The PDC composite sheet mold is placed on an ultrasonic scanning platform, and the surface of the PDC composite sheet mold is positioned and calibrated to obtain the mold scanning reference coordinates.

[0014] Based on the mold scanning reference coordinates, a multi-frequency ultrasonic probe scanning path is set, and the PDC composite mold is scanned point by point to obtain the ultrasonic reflection signal of each frequency channel.

[0015] The ultrasonic reflection signals of each frequency channel are normalized in amplitude and synchronized in the time domain to obtain standardized ultrasonic signal data.

[0016] The multi-channel original ultrasound image is obtained by performing pixel mapping and grayscale conversion processing on the standardized ultrasound signal data.

[0017] Optionally, step S2 includes:

[0018] The multi-channel original ultrasound image is input into a deep convolutional neural network, and image noise is filtered using a dual-scale convolutional kernel to obtain a denoised ultrasound image.

[0019] The denoised ultrasound image is subjected to adaptive histogram equalization processing, and the grayscale distribution is adjusted according to the acoustic impedance difference between diamond and cemented carbide materials to obtain a contrast-enhanced image.

[0020] The contrast-enhanced image is processed by combining the Laplacian and Sobel operators to enhance the image gradient, resulting in an edge-sharpened image.

[0021] The preprocessed ultrasound image is obtained by performing batch normalization and ReLU activation function processing on the edge-sharpened image.

[0022] Optionally, step S3 includes:

[0023] The preprocessed ultrasound image is input into the interface detection convolutional layer of the improved MesoNet algorithm, and the asymmetric convolutional kernel is used to perform linear feature extraction on the diamond-hard alloy interface to obtain the interface linear feature map.

[0024] Based on the interface linear feature map input multi-scale defect detection branch, small-scale, medium-scale, and large-scale convolution kernels are used to extract features from point defects, line defects, and area defects respectively, to obtain defect feature vectors at three scales.

[0025] Spatial attention weighting and feature concatenation are applied to the defect feature vectors at the three scales to obtain a comprehensive defect feature descriptor;

[0026] The comprehensive defect feature descriptor is processed by dimensional transformation and feature encoding through a fully connected layer to obtain the defect feature vector.

[0027] Optionally, the step of inputting the preprocessed ultrasound image into the interface detection convolutional layer of the improved MesoNet algorithm, and using the asymmetric convolutional kernel to perform linear feature extraction processing on the diamond-hard alloy interface to obtain an interface linear feature map, includes:

[0028] The preprocessed ultrasound images are channel aligned according to pixel positions to obtain multi-channel input data of uniform size.

[0029] Based on the multi-channel input data, 1×9 and 9×1 asymmetric convolution kernels are input respectively to perform convolution operations to obtain horizontal interface response maps and vertical interface response maps.

[0030] The horizontal and vertical interface response maps are fused pixel-by-pixel maximum values ​​to obtain a bidirectional interface response fusion map.

[0031] The bidirectional interface response fusion map is thresholded, and interface feature pixels above a set threshold are retained to obtain the interface linear feature map.

[0032] Optionally, step S4 includes:

[0033] The weight parameters of each frequency channel are calculated and processed based on the defect feature vector to obtain the fusion weight coefficient of each frequency channel image.

[0034] The multi-channel original ultrasound images are weighted and summed pixel by pixel according to the fusion weight coefficients to obtain the enhanced fused image;

[0035] The enhanced fused image input image segmentation algorithm is processed by convolution and pooling operations to obtain a downsampled multi-level encoded feature map;

[0036] Based on the multi-level encoded feature map, upsampling processing with transposed convolution and skip connections is performed to obtain a segmentation probability map. The segmentation probability map is then thresholded to obtain the binary defect segmentation mask.

[0037] Optionally, step S5 includes:

[0038] Based on the binary defect segmentation mask, pixel statistical processing is performed on independent defect regions to obtain pixel area data for each defect region.

[0039] The pixel area data of each defect region are summed to obtain the total defect pixel area.

[0040] The effective detection area of ​​the binary defect segmentation mask is processed by pixel counting to obtain the total pixel area of ​​the interface;

[0041] Based on the total defect pixel area and the total interface pixel area, a division operation is performed to obtain the percentage of defect area to the total interface area. Then, based on the comparison result of the percentage with the preset defect rate threshold, a quality grade classification is performed to obtain four quality grade judgment results for the PDC composite sheet mold: high-quality product, qualified product, substandard product, and scrap product.

[0042] Optionally, the step of performing pixel counting processing on the effective detection area of ​​the binary defect segmentation mask to obtain the total pixel area of ​​the interface includes:

[0043] The boundary contour extraction process is performed on the binary defect segmentation mask to obtain the outer boundary contour line of the PDC composite sheet mold;

[0044] Based on the outer boundary contour line, the internal region is filled to obtain the mold detection area mask;

[0045] The pixels with a value of 1 in the mask of the mold detection area are counted row by row and column by column to obtain the total number of valid detected pixels.

[0046] The total pixel area of ​​the interface is obtained by multiplying the total number of valid detected pixels by the actual area corresponding to a single pixel.

[0047] The technical solution provided in this application utilizes a multi-frequency ultrasonic probe to obtain multi-channel raw ultrasonic images, combined with a deep convolutional neural network (CNN) for denoising and contrast enhancement. This effectively improves image quality, enhances the grayscale distribution difference between the diamond and cemented carbide regions, and ensures that key features during the detection process are accurately captured. Especially at complex diamond-cemented carbide interfaces, the different penetration depths and resolutions of the multi-frequency probe allow for comprehensive information acquisition, avoiding the limitations of traditional single-frequency probes and improving the comprehensiveness and accuracy of the detection. Furthermore, through an improved MesoNet algorithm, asymmetric convolutional kernels are used to extract linear features at the interface, and multi-scale convolutional kernels are combined to detect different types of defects, enabling more detailed and accurate identification of various defect types, including point, line, and area defects. This technology not only optimizes the feature extraction process but also enhances the ability to identify diverse defects, making the quality assessment results more reliable.

[0048] The introduction of deep convolutional neural networks and the MesoNet algorithm has not only improved detection efficiency but also played a crucial role in the detection of defects at complex material interfaces. These algorithms' defect feature extraction and weighted fusion strategies at different scales effectively reduce background noise interference and improve defect detection accuracy. Particularly when processing the fusion of high-frequency and low-frequency information, the algorithms can weight the sensitivity of different frequency channels to defects, thereby ensuring optimal utilization of multi-channel data. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of one embodiment of the PDC composite sheet mold quality inspection method in this application;

[0051] Figure 2 This is a schematic diagram of the statistical results of the PDC composite sheet mold quality grade classification process in the embodiments of this application. Detailed Implementation

[0052] This application provides a method for quality inspection of PDC composite sheet molds. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the PDC composite sheet mold quality inspection method in this application includes:

[0054] Step S1: Obtain multi-channel raw ultrasound images by scanning the PDC composite mold with a multi-frequency ultrasound probe;

[0055] Step S2: Input the multi-channel raw ultrasound image into a deep convolutional neural network for denoising and contrast enhancement to generate a pre-processed ultrasound image, in which the diamond region and the cemented carbide region exhibit different grayscale value distributions.

[0056] Step S3: Feature extraction is performed on the preprocessed ultrasound image by improving the MesoNet algorithm. Asymmetric convolution kernels are used to extract linear features of the interface. Multi-scale convolution kernels are used to detect different types of defect features and output defect feature vectors.

[0057] Step S4: Establish a weighted fusion strategy based on the defect feature vector, perform pixel-level fusion of the multi-channel data corresponding to the preprocessed ultrasound image to generate an enhanced fused image, and obtain a binary defect segmentation mask through an image segmentation algorithm.

[0058] Step S5: Calculate the percentage of defect area to total interface area based on binary defect segmentation mask, and classify the PDC composite sheet mold quality into four quality levels: high-quality, qualified, substandard, and scrap according to the defect rate threshold.

[0059] It is understood that the executing entity of this application can be a PDC composite sheet mold quality inspection system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as an example for illustration.

[0060] Specifically, the multi-frequency ultrasonic probe simultaneously scans the PDC composite mold at four different frequencies: 25MHz, 50MHz, 75MHz, and 100MHz. The ultrasonic waves emitted by each frequency probe produce different degrees of reflection at the interface between the diamond layer and the hard alloy layer. The low-frequency probe has a large penetration depth but low resolution, while the high-frequency probe has high resolution but small penetration depth. By sampling and quantizing the received ultrasonic reflection signals according to the time sequence, the analog signal is converted into a digital signal. Then, the signal intensity is mapped to the corresponding pixel position according to the scanning position coordinates to form a 256×256 pixel grayscale image, where the pixel grayscale value reflects the reflection intensity of the ultrasonic wave at that position.

[0061] The deep convolutional neural network preprocessing stage filters the original ultrasound image using 3×3 and 5×5 dual-scale convolutional kernels. The 3×3 kernel detects local noise features, while the 5×5 kernel detects global noise patterns. The convolution operation removes high-frequency noise by multiplying the weight matrix with the image pixels. Adaptive histogram equalization adjusts the grayscale distribution based on the material properties of the PDC composite sheet, and diamond acoustic impedance... Corresponding to high grayscale values ​​of 180-220, hard alloy acoustic impedance For low grayscale values ​​of 60-100, pixel grayscale values ​​are redistributed by calculating the cumulative distribution function for each grayscale level. The Laplacian operator enhances edge gradients through second-order differentiation, while the Sobel operator detects edge directions through first-order differentiation in the horizontal and vertical directions. The combination of the two highlights defect boundary features.

[0062] The improved MesoNet algorithm employs asymmetric convolutional kernels to extract linear interface features. A 1×9 convolutional kernel specifically detects horizontal linear interfaces, while a 9×1 convolutional kernel detects vertical linear interfaces. Convolution operations are performed on the image using a sliding window, and the value at each location in the output feature map reflects the intensity of the interface feature at that location. In the multi-scale defect detection branch, the 3×3 convolutional kernel has a small receptive field, suitable for detecting point defects such as pores; the 5×5 convolutional kernel has a medium receptive field, suitable for detecting linear defects such as cracks; and the 7×7 convolutional kernel has a large receptive field, suitable for detecting planar defects such as delamination. Each branch outputs a 128-dimensional feature vector, which is concatenated to form a 384-dimensional comprehensive feature descriptor. The spatial attention mechanism calculates the importance weight of each location in the feature map, highlighting defect areas and suppressing background interference. The weight calculation is based on the variance and mean statistics of the feature values.

[0063] A weighted fusion strategy is established based on the defect feature vector. Different dimensions in the feature vector reflect the sensitivity of different frequency channels to defects. The feature values ​​are converted into weight coefficients using the softmax function, and the sum of the weight coefficients equals 1. The multi-channel preprocessed ultrasound image is then weighted and summed, that is, the gray values ​​of the four frequency channels at each pixel position are multiplied by their corresponding weight coefficients and then added to obtain the pixel value of the fused image. The encoder part of the image segmentation algorithm extracts features through convolutional layers and pooling layers reduce the resolution to form a multi-level feature representation. The decoder part upsamples through transposed convolution to restore the original resolution. Skip connections concatenate the features of corresponding levels from the encoder and decoder to retain detailed information, and finally output the probability value of each pixel belonging to a defect.

[0064] The segmentation probability map is thresholded with a threshold of 0.5. Pixels with a probability value greater than 0.5 are marked as defective pixels with a value of 1, and pixels with a probability value less than 0.5 are marked as normal pixels with a value of 0, forming a binary defect segmentation mask. Connectivity analysis uses an eight-neighbor search to group adjacent defective pixels into independent regions. The number of pixels in each region is counted to obtain the defect area. The total defect area is obtained by summing the areas of all defective regions. At the same time, the total number of pixels in the entire detection area is counted to obtain the total interface area. The defect rate percentage is calculated by dividing the two. Quality grade classification is determined according to preset thresholds: a defect rate of less than 2% is considered high-quality, 2%-5% is considered acceptable, 5%-10% is considered substandard, and greater than 10% is considered scrap.

[0065] In one specific embodiment, step S1 includes:

[0066] The PDC composite sheet mold was placed on an ultrasonic scanning platform, and the surface of the PDC composite sheet mold was positioned and calibrated to obtain the mold scanning reference coordinates.

[0067] The scanning path of the multi-frequency ultrasonic probe is set based on the mold scanning reference coordinates, and the PDC composite mold is scanned point by point to obtain the ultrasonic reflection signal of each frequency channel.

[0068] The ultrasonic reflection signals of each frequency channel are normalized in amplitude and synchronized in the time domain to obtain standardized ultrasonic signal data.

[0069] Pixel mapping and grayscale conversion are performed on standardized ultrasound signal data to obtain multi-channel raw ultrasound images.

[0070] Specifically, the positioning and calibration process of the PDC composite mold on the ultrasonic scanning platform is achieved by detecting the coordinates of three reference points on the mold surface using a laser displacement sensor. After the laser beam is perpendicularly irradiated onto the mold surface, it is reflected back to the sensor. The sensor calculates the distance value based on the optical path difference. The three reference points are located at the upper left, upper right, and lower left corners of the mold, respectively. The spatial position and rotation angle of the mold on the scanning platform are determined by these three points, and a two-dimensional coordinate system is established with the upper left corner as the origin. The X-axis points to the upper right corner, and the Y-axis points to the lower left corner. The mold scanning reference coordinates include the origin coordinates, the unit vector of the X-axis, and the unit vector of the Y-axis. These coordinate data directly determine the scanning trajectory and pixel mapping relationship of the subsequent ultrasonic probe.

[0071] The multi-frequency ultrasonic probe scanning path is set based on the scanning reference coordinates of the mold to calculate the position of the scanning grid points. The scanning path uses a rasterization method to divide the mold surface into a uniform grid. Each grid point corresponds to a pixel position in the subsequent image. The scanning step size is set to 0.1 mm to ensure sufficient spatial resolution. Four ultrasonic probes of different frequencies scan point by point along the same scanning path. Each probe emits an ultrasonic pulse after reaching the scanning point. When the ultrasonic wave propagates inside the PDC composite mold, it will be reflected when it encounters the interface between the diamond layer and the hard alloy layer. The intensity of the reflected wave is proportional to the difference in acoustic impedance at the interface. The probe receives the reflected ultrasonic signal and converts it into a voltage signal. The voltage amplitude reflects the ultrasonic reflection intensity. The four frequency probes obtain four different voltage values ​​at the same position. These voltage values ​​constitute the multi-frequency ultrasonic reflection signal at that position.

[0072] The amplitude normalization processing of the ultrasonic reflected signal converts the original voltage signal into a value with a uniform range. The maximum and minimum amplitudes of the signal for each frequency channel are calculated. Then, the amplitude of each sampling point is subtracted from the minimum value and divided by the difference between the maximum and minimum values ​​to obtain the normalized amplitude between 0 and 1. Time-domain synchronization processing solves the time delay problem between different frequency probes. Since there are slight differences in the physical positions of the four probes, the ultrasonic wave propagation time is slightly different. The time delay is determined by calculating the distance difference between each probe and the mold surface. The signal with the larger delay is shifted forward by the corresponding time amount so that the signals of the four frequency channels are aligned on the time axis. The standardized ultrasonic signal data contains the normalized amplitude sequence of the four frequency channels, and the length of each sequence is equal to the total number of scanning points.

[0073] Pixel mapping processing converts a one-dimensional scan point sequence into a two-dimensional image matrix. Based on the rasterization setting of the scan path, the i-th scan point corresponds to the pixel position in the i-th row and j-th column of the image. The row and column numbers are calculated from the position of the scan point in the grid. Grayscale conversion processing converts the normalized amplitude into a grayscale value between 0 and 255. The conversion formula is that the grayscale value is equal to the normalized amplitude multiplied by 255 and then rounded down. Four 256×256 pixel grayscale images are generated for each of the four frequency channels. The grayscale value of each pixel in the image directly reflects the reflection intensity of the ultrasound at that position. The multi-channel original ultrasound image is composed of these four grayscale images.

[0074] In one specific embodiment, step S2 includes:

[0075] The multi-channel raw ultrasound images are input into a deep convolutional neural network, and dual-scale convolutional kernels are used to filter the image noise to obtain a denoised ultrasound image.

[0076] Adaptive histogram equalization is performed on the denoised ultrasound image, and the grayscale distribution is adjusted according to the acoustic impedance difference between diamond and cemented carbide materials to obtain a contrast-enhanced image.

[0077] The contrast-enhanced image is processed by combining the Laplacian operator and the Sobel operator to enhance the image gradient, resulting in an edge-sharpened image.

[0078] Preprocessed ultrasound images are obtained by batch normalization and ReLU activation function processing based on edge-sharpened images.

[0079] Specifically, in the dual-scale convolutional kernel filtering process of the deep convolutional neural network inputting multi-channel raw ultrasound images, 3×3 and 5×5 convolutional kernels correspond to different filtering characteristics. The 3×3 convolutional kernel contains 9 weight parameters, which perform dot product operations with the pixel values ​​in each 3×3 neighborhood of the image. Specifically, the center of the convolutional kernel is aligned with the target pixel, and the 9 weight values ​​of the convolutional kernel are multiplied by the corresponding 9 pixel gray values ​​and then summed to obtain the filtered pixel value. This small-sized convolutional kernel mainly removes local high-frequency noise and salt-and-pepper noise. The 5×5 convolutional kernel contains 25 weight parameters, which perform convolution operations with the pixel values ​​in each 5×5 neighborhood of the image. The large-sized convolutional kernel has a larger receptive field and mainly removes global low-frequency noise and background interference. The weight parameters of the two types of convolutional kernels are learned through training data. During the training process, the network adjusts the weight values ​​according to the difference between the input noisy image and the corresponding clean image, so that the filtered image is closer to the clean image. The denoised ultrasound image is a weighted combination of the filtering results of the two types of convolutional kernels.

[0080] Adaptive histogram equalization adjusts the image's grayscale distribution based on the acoustic impedance difference between diamond and cemented carbide materials in the PDC composite sheet. The acoustic impedance of diamond is... The acoustic impedance of hard alloy is The acoustic impedance ratio between the two is approximately 0.63. This difference causes ultrasonic waves to reflect at different intensities at the interface. The diamond region exhibits a relatively high grayscale value in the ultrasonic image, while the cemented carbide region exhibits a relatively low grayscale value. Adaptive histogram equalization first statistically analyzes the pixel distribution of each grayscale level in the image, calculates the cumulative distribution function of each grayscale level, and then remaps the grayscale values ​​according to the target distribution, so that the grayscale values ​​of the diamond region are concentrated in the range of 180-220, and the grayscale values ​​of the cemented carbide region are concentrated in the range of 60-100, thereby enhancing the contrast between the two materials. With enhanced contrast, the grayscale jumps at the material interface in the image become more obvious.

[0081] Image gradient enhancement processing combining the Laplacian and Sobel operators targets the edge features of defects in PDC composite molds. The Laplacian operator is a second-order differential operator that detects edges in an image by calculating the second derivative of a pixel. Specifically, it subtracts the average gray value of the eight neighboring pixels from the gray value of the target pixel to obtain the Laplacian response value of that pixel. The larger the absolute value of the response value, the closer the pixel is to the edge. The Sobel operator is a first-order differential operator that calculates the gradients in the horizontal and vertical directions. The horizontal Sobel operator calculates the horizontal gradient based on the gray value difference between the left and right adjacent pixels, and the vertical Sobel operator calculates the vertical gradient based on the gray value difference between the top and bottom adjacent pixels. The square root of the sum of the squares of the gradients in both directions is used to obtain the gradient magnitude, which reflects the edge strength at that pixel. The Laplacian operator is sensitive to noise but can accurately locate edges, while the Sobel operator has strong noise resistance but slightly coarser edge localization. The combination of the two ensures both accurate edge localization and suppresses noise interference, resulting in clearer contours of defects in the edge-sharpened image.

[0082] Batch normalization standardizes the grayscale value of each pixel in the edge-sharpened image. First, it calculates the mean and variance of the pixel grayscale values ​​of the entire image. Then, it subtracts the mean from each pixel value and divides by the standard deviation to obtain a standardized data distribution with a mean of 0 and a variance of 1. This process eliminates the differences in brightness and contrast between different images. The ReLU activation function performs a non-linear transformation on the standardized pixel values. The mathematical expression of the ReLU function is that the output is equal to the maximum value of the input value and 0, that is, negative values ​​become 0 while positive values ​​remain unchanged. This process removes negative pixel values ​​while retaining positive feature responses, enhancing the expressive power of effective features in the image. After the above four steps of data processing, the preprocessed ultrasound image has noise suppressed, contrast enhanced, edge features sharpened, and data distribution standardized.

[0083] For example, in the original image of a 25MHz frequency channel of a PDC composite die, the pixel grayscale distribution of a certain 3×3 region is as follows: the central pixel is 120, and the surrounding 8 pixels are 118, 125, 116, 132, 108, 127, 119, and 124 respectively. After filtering with a 3×3 convolution kernel, the new value of the central pixel is the sum of the product of each pixel value and its corresponding convolution kernel weight. Assuming the convolution kernel weight is 0.5 for the center and 0.0625 for each of the surrounding pixels, the filtered pixel value is 120 multiplied by 0.5 plus the sum of the surrounding 8 pixel values ​​multiplied by 0.0625, resulting in 75.4375. A similar calculation is performed when using a 5×5 convolution kernel, considering 25 pixel values ​​in a larger neighborhood, for adaptive histogram equalization. The pixel's grayscale value of 120 is mapped to the corresponding new grayscale value in the target distribution. Based on the characteristics of diamond material, this pixel should belong to the diamond region. Mapping it to the range of 180-220 yields a new grayscale value of 195. The Laplacian operator calculates the second derivative response of this pixel as 120 minus the mean of the surrounding 8 pixels, 121.125, resulting in -1.125. The Sobel operator calculates the horizontal gradient as the mean of the right pixels minus the mean of the left pixels, and the vertical gradient as the mean of the lower pixels minus the mean of the upper pixels. The combined magnitude of the two gradients reflects the edge strength. Batch normalization subtracts the image mean from the processed pixel value and divides it by the standard deviation to obtain a standardized value. The ReLU activation function sets negative values ​​to zero and retains positive values, finally yielding the preprocessed pixel value.

[0084] In one specific embodiment, step S3 includes:

[0085] The preprocessed ultrasound image is input into the interface detection convolutional layer of the improved MesoNet algorithm. Asymmetric convolutional kernels are used to perform linear feature extraction on the diamond-hard alloy interface to obtain the interface linear feature map.

[0086] Based on the interface linear feature map input multi-scale defect detection branch, small-scale, medium-scale and large-scale convolution kernels are used to extract features of point defects, line defects and area defects respectively, to obtain defect feature vectors of three scales;

[0087] Spatial attention weighting and feature concatenation are applied to the defect feature vectors at three scales to obtain a comprehensive defect feature descriptor.

[0088] The comprehensive defect feature descriptor is processed by dimensional transformation and feature encoding through a fully connected layer to obtain the defect feature vector.

[0089] Specifically, when preprocessing the ultrasound image input into the interface detection convolutional layer of the improved MesoNet algorithm, the asymmetric convolutional kernel is specifically designed to extract the linear structural features of the diamond-hard alloy interface in the PDC composite mold. The 1×9 asymmetric convolutional kernel contains 9 weight parameters arranged in a horizontal line. When the convolutional kernel slides on the image, it mainly detects the linear features in the horizontal direction. Specifically, the kernel is convolved with each 1×9 neighborhood in the image, and the 9 weight values ​​are multiplied by the corresponding 9 pixel gray values ​​and then summed to obtain the horizontal linear feature response value at that position. The 9×1 asymmetric convolutional kernel contains 9 weight parameters arranged in a vertical line and detects the linear features in the vertical direction in the same way. Since the interface between the diamond layer and the hard alloy layer in the PDC composite mold usually exhibits a regular linear distribution, this asymmetric convolutional kernel structure can effectively capture the geometric features of the interface. The interface linear feature map is obtained by fusing the feature responses in the horizontal and vertical directions pixel by pixel. The fusion method is to take the maximum value of the response values ​​in the two directions to highlight the linear structural features of the interface region.

[0090] The multi-scale defect detection branch extracts features of different types of defects based on the interface linear feature map. The small-scale branch uses a 3×3 convolution kernel to specifically detect point defects in PDC composite molds. Point defects mainly include local defects such as pores and voids. These defects appear as isolated bright spots or dark spots in ultrasound images. The receptive field of the 3×3 convolution kernel precisely covers the spatial scale of a single point defect. The convolution operation extracts point features by calculating the weighted sum of the center pixel and its eight neighboring pixels. The medium-scale branch uses a 5×5 convolution kernel to detect linear defects, which mainly include extended defects such as cracks and microcracks. Defects appear as continuous linear dark or bright areas in ultrasound images. The receptive field of a 5×5 convolutional kernel can cover the width and part of the length of linear defects. The large-scale branch uses a 7×7 convolutional kernel to detect planar defects, which mainly include large-area defects such as delamination and cold welds. These defects appear as irregular sheet-like areas in ultrasound images. The large receptive field of the 7×7 convolutional kernel can capture the overall shape features of planar defects. The three scale convolutional kernels perform convolution operations on the interface linear feature map respectively. Each branch outputs a 128-dimensional feature vector. Each dimension of the feature vector corresponds to the feature intensity at a certain spatial location in the feature map after convolution.

[0091] Spatial attention weighted processing assesses the importance of defect feature vectors at three scales. The spatial attention mechanism determines the weights by calculating the statistical properties of each dimension in the feature vector. First, the mean and variance of each feature vector are calculated. The mean reflects the average activation intensity of the feature, and the variance reflects the dispersion of the feature distribution. Then, the mean and variance are input into a fully connected network to calculate the attention weights. The attention weights represent the importance of the corresponding feature dimension. The larger the weight value, the greater the contribution of that dimension to defect detection. The calculated attention weights are multiplied element-wise with the original feature vectors to obtain the weighted feature vectors. This processing method can highlight important feature dimensions while suppressing irrelevant feature dimensions. Feature concatenation processing connects the three weighted 128-dimensional feature vectors end to end to form a 384-dimensional comprehensive defect feature descriptor. The concatenation operation simply arranges the elements of the three vectors in order into a long vector, preserving the complete information of defect features at different scales.

[0092] The dimensional transformation and feature encoding processing of the fully connected layer convert the 384-dimensional comprehensive defect feature descriptor into the final defect feature vector. The fully connected layer contains a weight matrix and a bias vector. The weight matrix has a dimension of 384×256, and the bias vector has a dimension of 256. The dimensional transformation operation is to multiply the input 384-dimensional vector with the weight matrix and then add the bias vector to obtain a 256-dimensional intermediate feature vector. This linear transformation can learn the correlation and importance between different feature dimensions. The feature encoding processing performs a non-linear transformation on the intermediate feature vector through an activation function. The activation function used is the ELU function. The ELU function keeps positive values ​​unchanged and exponentially decays negative values. This processing method can retain important positive feature responses while smoothing negative feature responses. The final defect feature vector contains comprehensive feature information of all types of defects in the PDC composite die.

[0093] In one specific embodiment, the process of inputting the preprocessed ultrasound image into the interface detection convolutional layer of the improved MesoNet algorithm can specifically include the following steps:

[0094] The preprocessed ultrasound images are channel aligned according to pixel positions to obtain multi-channel input data of uniform size.

[0095] Based on the multi-channel input data, 1×9 and 9×1 asymmetric convolution kernels are input respectively to perform convolution operations to obtain the horizontal interface response map and the vertical interface response map;

[0096] The horizontal and vertical interface response maps are fused pixel by pixel to obtain a bidirectional interface response fusion map.

[0097] The bidirectional interface response fusion map is thresholded, and interface feature pixels above the set threshold are retained to obtain the interface linear feature map.

[0098] Specifically, the channel alignment processing of the preprocessed ultrasound images addresses the spatial position deviation problem caused by multi-frequency ultrasound probe acquisition. Due to slight positional differences in the physical installation of the four ultrasound probes of different frequencies, the pixel coordinates corresponding to the same scanning point in different frequency channels are slightly different. The channel alignment processing first calculates the spatial offset between the images of each frequency channel, and then finds the position of maximum correlation between the images of different channels through a cross-correlation algorithm. The cross-correlation operation uses one image as a template to slide on another image and calculates the correlation coefficient at each position. The position with the largest correlation coefficient corresponds to the best matching position, and the offset is the difference between this position and the original position. Then, based on the calculated offset, the images of each frequency channel are spatially transformed to align the pixel positions of all channels to a unified coordinate system, ensuring that the same physical position corresponds to the same pixel coordinates in all frequency channels. The multi-channel input data of uniform size includes aligned images of four frequency channels, with each channel image size being 256×256 pixels, and the spatial correspondence between pixels is completely consistent.

[0099] The 1×9 asymmetric convolution kernel is specifically designed for feature detection of the horizontal linear structure of the diamond-hard alloy interface in PDC composite die. The 1×9 convolution kernel contains 9 horizontally arranged weight parameters. During the convolution operation, the kernel slides pixel by pixel on the multi-channel input data. For each target pixel position, the kernel is multiplied by the 1×9 neighborhood around that position. Specifically, the 9 weight values ​​of the kernel are multiplied by the gray values ​​of the 9 pixels at the corresponding positions and then summed to obtain the horizontal feature response value at that position. Since the interface in the PDC composite die is usually horizontal or near-horizontal, the 1×9 convolution kernel can effectively capture this horizontal linear feature. The horizontal interface response map records the horizontal linear feature intensity of each pixel position. The 9×1 asymmetric convolution kernel uses the same convolution operation principle to detect the vertical linear features. The 9 weight parameters are arranged vertically and mainly detect the vertical components and boundary contours of the interface. The vertical interface response map records the vertical linear feature intensity of each pixel position. The response maps of the two directions together describe the complete geometric features of the interface in the PDC composite die.

[0100] The pixel-wise maximum value fusion processing merges the horizontal and vertical interface response maps into a single feature representation. The fusion operation compares the feature values ​​of corresponding pixel positions in the two response maps one by one, and selects the larger feature value as the fusion result. Mathematically, the value of each pixel in the fused image is equal to the maximum value of the corresponding pixel in the horizontal and vertical response maps. This fusion method can retain the stronger feature response in both directions, highlight the linear structural features of the interface area, and suppress background noise. The bidirectional interface response fusion map integrates the interface feature information in both the horizontal and vertical directions. The area with higher pixel value corresponds to the position of the diamond-hard alloy interface in the PDC composite sheet mold, and the area with lower pixel value corresponds to the position inside the uniform material. The fused image maintains the same 256×256 pixel size as the original image.

[0101] Thresholding converts the bidirectional interface response fusion map into a binary feature map by setting a fixed threshold. The threshold is selected based on the statistical distribution of feature response values ​​in the fusion image, usually set to the upper quartile of the pixel value distribution of the fusion image, i.e., 75% of the pixel values ​​are less than the threshold and 25% of the pixel values ​​are greater than the threshold. The thresholding operation compares the feature response value of each pixel with the set threshold. Pixels greater than the threshold are marked as interface feature pixels and assigned a value of 1, while pixels less than or equal to the threshold are marked as background pixels and assigned a value of 0. This binarization process can effectively separate interface features and background areas. The interface linear feature map clearly identifies the precise location and geometry of the diamond-hard alloy interface in the PDC composite die in binary form. Pixels with a value of 1 in the binary image form a continuous linear structure, accurately reflecting the direction and integrity of the interface.

[0102] In one specific embodiment, step S4 includes:

[0103] The weight parameters of each frequency channel are calculated and processed based on the defect feature vector to obtain the fusion weight coefficient of each frequency channel image.

[0104] The original multi-channel ultrasound images are weighted and summed pixel by pixel according to the fusion weight coefficient to obtain the enhanced fused image;

[0105] The enhanced fused image input image segmentation algorithm is processed by convolution and pooling operations to obtain a downsampled multi-level encoded feature map;

[0106] Upsampling processing based on transposed convolution and skip connections of multi-level encoded feature maps is performed to obtain a segmentation probability map. The segmentation probability map is then thresholded to obtain a binary defect segmentation mask.

[0107] Specifically, the weighting parameter calculation of the defect feature vector is based on the analysis of the differences in defect sensitivity of different dimensions in the 256-dimensional defect feature vector for each frequency channel. The first 64 dimensions of the defect feature vector mainly reflect the deep, large-size defect features detected in the 25MHz low-frequency channel; dimensions 65-128 reflect the medium-sized interface defect features detected in the 50MHz mid-frequency channel; dimensions 129-192 reflect the shallow, fine defect features detected in the 75MHz high-frequency channel; and dimensions 193-256 reflect the surface micro-defect features detected in the 100MHz ultra-high-frequency channel. The weighting parameter calculation first assigns weights to the 64 dimensions corresponding to each frequency band. Statistical analysis is performed on the eigenvectors, and the L2 norm of each segment of the eigenvector is calculated as the feature intensity index of that frequency channel. The L2 norm is calculated as the square root of the sum of the squares of 64 eigenvalues. Then, the feature intensity indices of the four frequency channels are input into the softmax function for normalization. The softmax function converts four arbitrary real values ​​into a probability distribution that sums to 1. Specifically, the feature intensity index of each channel is calculated by taking the exponent and dividing it by the sum of the exponents of the four channels. The four values ​​obtained are the fusion weight coefficients of the images of each frequency channel. The weight coefficients reflect the relative importance of each frequency channel to the current PDC composite mold defect detection.

[0108] The pixel-wise weighted summation processing of multi-channel raw ultrasound images linearly combines the images of the four frequency channels according to the calculated fusion weight coefficients. The weighted summation operation is performed separately for each pixel position in the image. For the pixel position in the i-th row and j-th column of the image, the pixel gray value of each of the four frequency channels at that position is extracted. Then, the pixel value of each channel is multiplied by the corresponding fusion weight coefficient. Finally, the four weighted results are added together to obtain the fused pixel value. This pixel-wise processing method ensures that the enhanced fused image maintains the same spatial resolution and size as the original image. The gray value of each pixel in the enhanced fused image integrates the information of the four frequency channels. The frequency channels with larger weight coefficients contribute more to the fusion result, while the frequency channels with smaller weight coefficients play an auxiliary and supplementary role. The fused image highlights the frequency information most sensitive to the current defect type while retaining useful information from other frequencies.

[0109] The convolution and pooling operations in image segmentation algorithms transform the enhanced fused image into a multi-level feature representation. The convolution operation uses multiple convolution kernels of different sizes to extract features from the fused image. The first convolution layer uses 32 3×3 convolution kernels to detect local edge and texture features. Each convolution kernel slides across the image to perform convolution operations, outputting 32 feature maps. The second convolution layer uses 64 3×3 convolution kernels to perform higher-level feature extraction on the output of the first layer. The third convolution layer uses 128 3×3 convolution kernels to further abstract the feature representation. The pooling operation is performed after each convolution layer, using a 2×2 max pooling window to downsample the feature map. The pooling operation selects the maximum pixel value in each 2×2 region as the output, while reducing the spatial size of the feature map by half. The multi-level encoded feature map contains feature representations with different levels of abstraction and spatial resolution. Shallow feature maps retain detailed spatial information but have less semantic information, while deep feature maps have rich semantic information but lower spatial resolution.

[0110] The upsampling process using transposed convolution and skip connections restores the multi-layered encoded feature maps to the spatial resolution of the original image. Transposed convolution is the inverse of convolution, achieving upsampling by inserting zero values ​​between pixels in the input feature map and then performing a standard convolution operation. The first transposed convolution upsamples the deepest 128-channel feature map from 64×64 to 128×128, and the second transposed convolution upsamples the feature map from 128×128 to 256×256. Skip connections concatenate and fuse the feature maps of the corresponding layers in the encoding stage with those in the decoding stage. Specifically, this involves upsampling the 64-channel feature maps of the second layer of the encoder... The feature map and the feature map after upsampling in the first layer of the decoder are concatenated in the channel dimension to obtain a 192-channel fused feature map. This skip connection mechanism can pass shallow detailed information to deep semantic representation. The segmentation probability map is generated by the last 1×1 convolution, which converts the multi-channel feature map into a single-channel pixel classification probability. The value of each pixel represents the probability value of the pixel belonging to the defect. The thresholding process sets a fixed threshold, usually 0.5, and converts the probability map into a binary defect segmentation mask. Pixels with a probability value greater than the threshold are marked as defect pixels with a value of 1, and pixels with a probability value less than the threshold are marked as normal pixels with a value of 0.

[0111] In one specific embodiment, step S5 includes:

[0112] Pixel statistical processing is performed on independent defect regions based on binary defect segmentation mask to obtain pixel area data of each defect region.

[0113] The pixel area data of each defect region are summed to obtain the total defect pixel area.

[0114] Pixel counting is performed on the effective detection area of ​​the binary defect segmentation mask to obtain the total pixel area of ​​the interface.

[0115] The total defect pixel area and the total interface pixel area are divided to obtain the percentage of the defect area to the total interface area. The quality grade is then classified according to the percentage and the preset defect rate threshold, resulting in four quality grade judgments for the PDC composite sheet mold: high-quality, qualified, substandard, and scrap.

[0116] Specifically, the pixel statistical processing of independent defect regions in the binary defect segmentation mask uses a connected component analysis algorithm to identify and quantify each separated defect region. The connected component analysis algorithm scans the entire binary image using an eight-neighbor search method, checking each pixel row by row and column by column starting from the top left corner. When a defect pixel with a value of 1 is encountered, the algorithm starts a depth-first search or breadth-first search strategy, recursively checking the eight neighboring pixels of that pixel. If the value of a neighboring pixel is also 1 and has not been visited, it is marked as the same connected component, and the search for its neighbors continues. This search process continues until all pixels in the connected component are found. Then, a unique identifier is assigned to the connected component, and the search continues for the next unmarked defect pixel, repeating the above process. Finally, all independent defect regions in the PDC composite sheet mold are identified. Each connected component represents an independent defect region. Pixel statistical processing calculates the number of pixels contained in each connected component, that is, the pixel area data of the defect region. These data reflect the spatial size of different defects.

[0117] The cumulative summation of pixel area data for each defect region involves summing the pixel counts of all independent defect regions. The summation operation iterates through all defect region identifiers in the connected component analysis results, adding the pixel area data of each region. The summation process involves initializing the total variable to zero, then sequentially reading the pixel count of the first defect region and adding it to the total, then reading the pixel count of the second defect region and adding it to the total, and so on until all defect regions have been processed. The total defect pixel area is the arithmetic sum of the pixel counts of all independent defect regions. This value reflects the overall scale of defects in the PDC composite die. Regardless of whether the defects are concentrated or dispersed, the cumulative summation can accurately calculate the total defect area.

[0118] The pixel counting process of the effective detection area of ​​the binary defect segmentation mask determines the total area range of the PDC composite mold interface. The effective detection area refers to the area where interface information can be effectively detected during ultrasonic scanning. This area excludes invalid areas at the edge of the mold and areas where ultrasonic waves cannot penetrate. The pixel counting process first performs morphological processing on the binary defect segmentation mask, filling small voids and connecting broken boundaries through expansion and erosion operations. Then, the outer contour of the mold is extracted using an edge detection algorithm. The edge detection algorithm scans the entire binary image to find the positions where the pixel value jumps from 0 to 1 or from 1 to 0. These positions constitute the boundary contour of the mold. All pixels inside the contour constitute the effective detection area. The pixel counting operation traverses every pixel position inside the contour and counts the total number of pixels to obtain the total pixel area of ​​the interface. This area data represents the complete range of the diamond-hard alloy interface in the PDC composite mold.

[0119] The division operation calculates the defect rate percentage by dividing the total defective pixel area by the total interface pixel area. The dividend in the division operation is the total defective pixel area calculated in the previous steps, and the divisor is the total interface pixel area of ​​the effective detection area. The quotient is the proportion of defective pixels to the total pixels. This proportion is multiplied by 100 to obtain the defect rate value in percentage form. The quality grade classification process determines the grade based on the comparison result of the defect rate percentage and the preset threshold. The preset defect rate threshold is determined according to the quality standards and usage requirements of the PDC composite sheet mold. Usually, multiple thresholds are set to divide the quality into four grades. The comparison process uses conditional judgment logic. If the defect rate is less than the first threshold, it is judged as a high-quality product. If the defect rate is between the first and second thresholds, it is judged as a qualified product. If the defect rate is between the second and third thresholds, it is judged as a substandard product. If the defect rate is greater than the third threshold, it is judged as a scrap product. The quality grade determination result of the PDC composite sheet mold directly corresponds to its applicability and reliability in practical applications.

[0120] Figure 2This diagram illustrates the statistical results of the quality grade classification of PDC composite sheet molds in this embodiment. The diagram shows the quality grade distribution after inspecting 300 PDC composite sheet mold samples using the method of this invention. Based on the comparison between the percentage of defect area to the total interface area and a preset defect rate threshold, the samples are divided into four quality grades: 156 high-quality products (defect rate <2%, accounting for 52.0%), 89 qualified products (defect rate 2%-5%, accounting for 29.7%), 45 substandard products (defect rate 5%-10%, accounting for 15.0%), and 10 scrap products (defect rate >10%, accounting for 3.3%). This distribution verifies the effectiveness of the quality grade classification process. The combined proportion of high-quality and qualified products reaches 81.7%, indicating that the detection method of this invention can accurately identify high-quality PDC composite sheet molds, providing a reliable basis for quality control in actual production. The different grayscale bars in the diagram create a clear visual contrast, intuitively reflecting the technical effect of quality classification based on binary defect segmentation masks.

[0121] In one specific embodiment, the process of performing pixel counting processing on the effective detection area of ​​the binary defect segmentation mask can specifically include the following steps:

[0122] The boundary contour of the PDC composite sheet mold is obtained by performing boundary contour extraction on the binary defect segmentation mask.

[0123] The mold detection area mask is obtained by filling the internal region based on the outer boundary contour line.

[0124] The pixels with a value of 1 in the mold detection area mask are counted row by row and column by column to obtain the total number of valid detected pixels.

[0125] The total pixel area of ​​the interface is obtained by multiplying the total number of effectively detected pixels by the actual area corresponding to a single pixel.

[0126] Specifically, the boundary contour extraction process of the binary defect segmentation mask uses the Canny edge detection algorithm to identify the outer boundary contour of the PDC composite mold. The Canny algorithm first performs Gaussian filtering on the binary mask image to remove noise, and then calculates the gradient magnitude and gradient direction of each pixel in the image. The gradient magnitude is calculated by the horizontal and vertical convolution results of the Sobel operator, and the gradient direction is the arctangent of the horizontal gradient and the vertical gradient. Non-maximum suppression processing checks whether the gradient magnitude of each pixel is a local maximum along the gradient direction. If the gradient magnitude of a pixel is less than the magnitude of its neighboring pixels in the gradient direction, it is set to zero. Dual threshold detection sets a high threshold and a low threshold to classify the gradient magnitude. Pixels above the high threshold are marked as strong edges, pixels below the low threshold are marked as non-edges, and pixels between the two thresholds are marked as weak edges. Edge connection processing promotes weak edge pixels connected to strong edges to strong edges through 8-neighborhood search. The final outer boundary contour completely describes the geometric boundary of the PDC composite mold in the ultrasound image. The contour is composed of a series of continuous pixels, and the coordinates of these pixels constitute the outer contour data of the mold.

[0127] The internal region filling process marks all pixel regions inside the mold as valid detection areas based on the outer boundary contour. The filling algorithm uses a scan-line seed filling method to achieve region filling. The seed filling algorithm first selects any point inside the contour as a seed point, and then expands the search in four directions from the seed point, marking all pixels connected to the seed point and located inside the contour as filling pixels. The scan-line optimized seed filling accelerates the filling process through horizontal scan lines. The algorithm fills continuous pixel segments along the horizontal direction, and then checks whether there are pixels that need to be filled in the previous and next rows. If they exist, they are used as new seed points to continue filling. The filling process continues until all pixels inside the contour are marked. The mold detection area mask is a binary image of the same size as the original image, where the area with a pixel value of 1 represents the effective detection range of the PDC composite mold, and the area with a pixel value of 0 represents the invalid area outside the mold. This mask image clearly defines the effective range of quality detection.

[0128] The row-by-row, column-by-column traversal counting process performs statistical calculations on all pixels with a value of 1 in the mask of the mold detection area. The traversal algorithm uses a double loop structure to implement pixel counting. The outer loop controls the row index of the image to traverse from the first row to the last row, and the inner loop controls the column index of the image to traverse from the first column to the last column. For each pixel position, the algorithm reads the pixel value at that position and determines whether it is equal to 1. If the pixel value is 1, the counter is incremented by 1. If the pixel value is 0, the pixel is skipped and the next position is continued. The initial value of the counter is set to 0. After the traversal is completed, the final value of the counter is the total number of valid detected pixels. This value reflects the total number of pixels of the PDC composite mold in the ultrasonic scanning image. The traversal process is carried out in the order of left to right and top to bottom to ensure that each pixel position is checked and checked only once, avoiding duplicate counting or omissions.

[0129] The area conversion process transforms the total number of effectively detected pixels into an actual physical area value. The conversion calculation is based on the spatial resolution parameter of the ultrasonic scan, which is determined by the scanning step length. The scanning step length is the minimum distance interval at which the ultrasonic probe moves on the surface of the PDC composite mold. Each pixel corresponds to a small square area on the mold surface, and the side length of this square area is equal to the scanning step length. The actual area corresponding to a single pixel is equal to the square of the scanning step length. The conversion formula is that the total pixel area of ​​the interface is equal to the total number of effectively detected pixels multiplied by the actual area of ​​a single pixel. This conversion process converts the pixel count in the image domain into an area measurement in the physical domain. The unit of the total pixel area of ​​the interface is square millimeters. This value accurately reflects the actual physical area size of the diamond-hard alloy interface in the PDC composite mold. The conversion result is directly used for subsequent defect rate calculation and quality grade determination.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quality inspection of PDC composite sheet molds, characterized in that, The method comprises: Step S1: Obtain multi-channel raw ultrasound images by scanning the PDC composite mold with a multi-frequency ultrasound probe; Step S2: Input the multi-channel original ultrasound image into a deep convolutional neural network for noise reduction and contrast enhancement to generate a preprocessed ultrasound image, wherein the diamond region and the cemented carbide region exhibit different grayscale value distributions. Step S3: The preprocessed ultrasound image is feature extracted by improving the MesoNet algorithm. Asymmetric convolution kernels are used to extract linear features of the interface, and multi-scale convolution kernels are used to detect different types of defect features, and the defect feature vector is output. Step S4: Establish a weighted fusion strategy based on the defect feature vector, perform pixel-level fusion of the multi-channel data corresponding to the preprocessed ultrasound image to generate an enhanced fusion image, and obtain a binary defect segmentation mask through an image segmentation algorithm. Step S5: Calculate the percentage of defect area to total interface area based on the binary defect segmentation mask, and classify the PDC composite sheet mold quality into four quality levels: high-quality, qualified, substandard, and scrap, according to the defect rate threshold.

2. The method for quality inspection of PDC composite sheet molds according to claim 1, characterized in that, Step S1 includes: The PDC composite sheet mold is placed on an ultrasonic scanning platform, and the surface of the PDC composite sheet mold is positioned and calibrated to obtain the mold scanning reference coordinates. Based on the mold scanning reference coordinates, a multi-frequency ultrasonic probe scanning path is set, and the PDC composite mold is scanned point by point to obtain the ultrasonic reflection signal of each frequency channel. The ultrasonic reflection signals of each frequency channel are normalized in amplitude and synchronized in the time domain to obtain standardized ultrasonic signal data. The multi-channel original ultrasound image is obtained by performing pixel mapping and grayscale conversion processing on the standardized ultrasound signal data.

3. The method for quality inspection of PDC composite sheet molds according to claim 1, characterized in that, Step S2 includes: The multi-channel original ultrasound image is input into a deep convolutional neural network, and image noise is filtered using a dual-scale convolutional kernel to obtain a denoised ultrasound image. The denoised ultrasound image is subjected to adaptive histogram equalization processing, and the grayscale distribution is adjusted according to the acoustic impedance difference between diamond and cemented carbide materials to obtain a contrast-enhanced image. The contrast-enhanced image is processed by combining the Laplacian and Sobel operators to enhance the image gradient, resulting in an edge-sharpened image. The preprocessed ultrasound image is obtained by performing batch normalization and ReLU activation function processing on the edge-sharpened image.

4. The method for quality inspection of PDC composite sheet molds according to claim 1, characterized in that, Step S3 includes: The preprocessed ultrasound image is input into the interface detection convolutional layer of the improved MesoNet algorithm, and the asymmetric convolutional kernel is used to perform linear feature extraction on the diamond-hard alloy interface to obtain the interface linear feature map. Based on the interface linear feature map input multi-scale defect detection branch, small-scale, medium-scale, and large-scale convolution kernels are used to extract features from point defects, line defects, and area defects respectively, to obtain defect feature vectors at three scales. Spatial attention weighting and feature concatenation are applied to the defect feature vectors at the three scales to obtain a comprehensive defect feature descriptor; The comprehensive defect feature descriptor is processed by dimensional transformation and feature encoding through a fully connected layer to obtain the defect feature vector.

5. The method for quality inspection of PDC composite sheet molds according to claim 4, characterized in that, The preprocessed ultrasound image is input into the interface detection convolutional layer of the improved MesoNet algorithm, and the asymmetric convolutional kernel is used to perform linear feature extraction on the diamond-hard alloy interface to obtain an interface linear feature map, including: The preprocessed ultrasound images are channel aligned according to pixel positions to obtain multi-channel input data of uniform size. Based on the multi-channel input data, 1×9 and 9×1 asymmetric convolution kernels are input respectively to perform convolution operations to obtain horizontal interface response maps and vertical interface response maps. The horizontal and vertical interface response maps are fused pixel-by-pixel maximum values ​​to obtain a bidirectional interface response fusion map. The bidirectional interface response fusion map is thresholded, and interface feature pixels above a set threshold are retained to obtain the interface linear feature map.

6. The method for quality inspection of PDC composite sheet molds according to claim 1, characterized in that, Step S4 includes: The weight parameters of each frequency channel are calculated and processed based on the defect feature vector to obtain the fusion weight coefficient of each frequency channel image. The multi-channel original ultrasound images are weighted and summed pixel by pixel according to the fusion weight coefficients to obtain the enhanced fused image; The enhanced fused image input image segmentation algorithm is processed by convolution and pooling operations to obtain a downsampled multi-level encoded feature map; Based on the multi-level encoded feature map, upsampling processing with transposed convolution and skip connections is performed to obtain a segmentation probability map. The segmentation probability map is then thresholded to obtain the binary defect segmentation mask.

7. The method for quality inspection of PDC composite sheet molds according to claim 1, characterized in that, Step S5 includes: Based on the binary defect segmentation mask, pixel statistical processing is performed on independent defect regions to obtain pixel area data for each defect region. The pixel area data of each defect region are summed to obtain the total defect pixel area. The effective detection area of ​​the binary defect segmentation mask is processed by pixel counting to obtain the total pixel area of ​​the interface; Based on the total defect pixel area and the total interface pixel area, a division operation is performed to obtain the percentage of defect area to the total interface area. Then, based on the comparison result of the percentage with the preset defect rate threshold, a quality grade classification is performed to obtain four quality grade judgment results for the PDC composite sheet mold: high-quality product, qualified product, substandard product, and scrap product.

8. The method for quality inspection of PDC composite sheet molds according to claim 7, characterized in that, The step of performing pixel counting processing on the effective detection area of ​​the binary defect segmentation mask to obtain the total pixel area of ​​the interface includes: The boundary contour extraction process is performed on the binary defect segmentation mask to obtain the outer boundary contour line of the PDC composite sheet mold; Based on the outer boundary contour line, the internal region is filled to obtain the mold detection area mask; The pixels with a value of 1 in the mask of the mold detection area are counted row by row and column by column to obtain the total number of valid detected pixels. The total pixel area of ​​the interface is obtained by multiplying the total number of valid detected pixels by the actual area corresponding to a single pixel.

Images