PHEMT defect detection method and system

Abstract

The invention discloses a PHEMT defect detection method and system, and relates to the technical field of electronic device detection, and the method comprises the steps: obtaining a PHEMT surface image; preprocessing the surface image to determine a target image; calculating a gradient response image of the target image; performing iterative threshold segmentation on the gradient response image to determine a final mask gradient image; according to the target image, through Gaussian difference and fixed threshold segmentation, determining a Gaussian segmentation image; combining the Gaussian segmentation image and the final mask gradient image, and determining a defect area; extracting defect features of the defect area; and according to the defect features, performing defect classification through a deep residual convolutional network, and outputting the defect category of the PHEMT. According to the invention, the defect area can be determined more accurately, the problem that the defect area is incomplete is avoided, the accuracy of defect area identification is improved, and the integrity and accuracy of the defect area are ensured.

Description

Technical Field

[0001] This invention relates to the field of electronic device testing technology, and in particular to a defect detection method and system for PHEMT. Background Technology

[0002] PHEMTs are semiconductor devices based on high electron mobility technology, widely used in high-frequency, high-power electronic devices such as radar, communication, and satellite technologies. Due to their complex manufacturing processes and harsh operating environments, PHEMTs are prone to various defects during production or use, such as cracks, surface scratches, material inhomogeneities, bubbles, or inclusions. These defects can lead to performance degradation or even device failure; therefore, efficient defect detection technology is crucial to ensuring the quality and reliability of PHEMTs. Currently, PHEMT defect detection methods typically combine modern image processing techniques and deep learning models, such as hyperspectral image analysis, gradient response image processing, differential Gaussian image processing, and deep residual convolutional networks (DFFNs), to achieve automated defect identification and classification. However, traditional defect detection methods still have many shortcomings. First, traditional methods often rely on manually designed features, making it difficult to fully extract effective features for complex defect morphologies or minute defects, resulting in low detection accuracy and difficulty adapting to diverse defect types and complex detection environments. Second, traditional defect classification methods are usually based on simple rules, unable to handle complex defect categories, often resulting in low classification accuracy and affecting the overall reliability and accuracy of the detection. Furthermore, traditional methods are susceptible to interference from uneven illumination, background noise, and complex textures during image segmentation, leading to incomplete defect region extraction or blurred boundaries, further reducing detection accuracy. To address these issues, a technical solution is urgently needed that can accurately extract defect regions, effectively distinguish complex defect categories, and improve detection precision. This invention combines gradient response images, adaptive thresholding, Gaussian difference, and deep residual convolutional networks to achieve more accurate defect region localization and classification, thereby overcoming the limitations of existing technologies and improving the overall performance and reliability of PHEMT defect detection. Summary of the Invention

[0003] This invention addresses the shortcomings of existing PHEMT defect detection methods, such as reliance on manually designed features, difficulty in adapting to complex defect morphologies or minute defects, and low classification accuracy. Therefore, this invention adopts the following technical solution: This invention provides a defect detection method for PHEMT (Problem Detection Technology), aiming to achieve accurate segmentation and classification of defect regions by combining image processing technology and deep learning models. The method includes the following steps: S1: Obtain the surface image of PHEMT.

[0004] High-resolution imaging equipment is used to acquire images of the surface of the PHEMT device to ensure image clarity and detail integrity. The high-resolution imaging equipment features uniform illumination compensation to reduce image quality fluctuations caused by changes in ambient light.

[0005] S2: Preprocess the surface image to determine the target image.

[0006] Furthermore, the preprocessing includes normalization and flattening. The normalization process performs a linear transformation on the image pixel values, adjusting the pixel value range to the [0,1] interval to eliminate the influence of uneven illumination. The flattening process uses a background modeling algorithm to remove background inhomogeneities in the image, highlighting the feature information of defective areas.

[0007] S3: Calculate the gradient response image of the target image.

[0008] The Sobel operator is used to perform convolution operations on the target image to extract regions with significant brightness variations. The Sobel operator calculates the gradient magnitude in both the horizontal and vertical directions, and obtains the gradient intensity for each pixel by summing the squares and taking the square root. This gradient response image is used to locate defect edges, providing foundational data for subsequent thresholding.

[0009] S4: Perform iterative threshold segmentation on the gradient response image to determine the final mask gradient image.

[0010] Furthermore, the iterative thresholding segmentation includes adaptive thresholding segmentation and mask optimization. The adaptive thresholding segmentation dynamically adjusts the segmentation threshold based on local grayscale information, specifically by constructing a Gaussian weight matrix to perform a weighted average of the gradient response image, generating a binary image. The kernel function of the Gaussian weight matrix is ​​defined as... in s Let x and y be the standard deviation of the Gaussian kernel, and x and y be the spatial coordinates of the kernel center. The mask optimization involves inverting the binary image and using it as a mask to remove strong gradient regions from the gradient response image while retaining low-contrast regions related to defects. The above steps are repeated until the iteration termination condition is met, generating the final mask gradient image. The iteration termination condition is that the change in mask gradient information between two consecutive iterations is less than a preset threshold. dori And the number of iterations did not exceed the maximum number of iterations N.

[0011] S5: Based on the target image, determine the Gaussian segmented image by using Gaussian difference and fixed threshold segmentation.

[0012] Furthermore, the Gaussian difference method obtains multi-scale feature information by constructing a scale-space image. The formula for generating the scale-space image is as follows: The Gaussian difference formula is as follows: The Gaussian difference result is then subjected to fixed threshold segmentation to generate a Gaussian segmented image. This fixed threshold segmentation uses a global threshold Tg to mark pixels in the Gaussian difference image larger than Tg as defect regions, and the remaining pixels as background regions.

[0013] S6: Merge the Gaussian segmentation image and the final mask gradient image to determine the defect region.

[0014] Further, it is determined whether the Gaussian segmentation image and the final mask gradient image can be merged. This is done by calculating the spatial distance d between them and determining if it is less than a preset spatial distance threshold ds. The spatial distance formula is as follows: ,in p x and p y These are the corresponding pixel coordinates of the Gaussian segmentation image and the final mask gradient image, respectively. If the merging condition is met, the two images are fused using a logical AND operation to generate the complete defect region.

[0015] S7: Extract the defect features of the defect region.

[0016] Furthermore, texture, shape, and color features are extracted from the defect area. The texture features are calculated using a gray-level co-occurrence matrix (GLCM) and include contrast, correlation, and energy indices. The shape features are obtained using a boundary contour fitting algorithm and include area, perimeter, and roundness. The color features are obtained through statistical analysis of the RGB channels of the defect area, including mean, variance, and skewness.

[0017] S8: Based on the defect characteristics, perform defect classification using a deep residual convolutional network.

[0018] Furthermore, the deep residual convolutional network includes an input layer, a residual layer, a feature fusion layer, a fully connected layer, a softmax classification layer, and an output layer. The residual layer consists of multiple convolutional units, batch normalization units, and a ReLU activation function. The feature fusion layer integrates multi-scale features through a concatenation operation to generate a fused feature vector. The fully connected layer maps the fused feature vector to a high-dimensional space to generate the feature representation required for classification. The output layer determines the final defect category label through an arg max operation.

[0019] S9: Optimize the deep residual convolutional network.

[0020] Furthermore, the Seagull Optimization Algorithm (SOA) is used to optimize the network weights. The positions of the seagull flock are initialized, with each position corresponding to a set of weight parameters for the network. The cross-entropy loss function is used as the fitness function to calculate the fitness value for each seagull. The positions of the seagulls are updated based on their fitness values, incorporating the motion behavior coefficient V and the global search performance U of the equilibrium algorithm. The optimization process is repeated until the maximum number of iterations is reached to determine the optimal weight parameters.

[0021] The beneficial effects of the present invention are: By combining gradient response images and iterative thresholding with Gaussian difference and fixed thresholding, defect regions can be effectively extracted, reducing background noise interference and improving segmentation accuracy. Deep residual convolutional networks automatically learn defect features, avoiding the limitations of manually designed features, and can distinguish complex defect categories, significantly improving classification accuracy. The Seagull optimization algorithm optimizes network weights, improving model training efficiency and generalization ability, ensuring system reliability in different application scenarios. The entire detection process, from image acquisition to defect classification, is fully automated, reducing manual intervention and improving detection efficiency.

[0022] The present invention also provides a defect detection system for PHEMT, including a processor and a memory. The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement all the steps of the defect detection method described above.

[0023] Specifically, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or an application-specific integrated circuit (ASIC), etc., and the memory may be volatile memory or non-volatile memory, including random access memory (RAM), read-only memory (ROM), or flash memory, etc.

[0024] The present invention also provides a readable storage medium having a program or instructions stored thereon that, when executed by a processor, implements all the steps of the above-described PHEMT defect detection method.

[0025] In summary, this invention, by combining image processing technology and deep learning models, achieves high-precision detection and classification of PHEMT defects, solves the technical problems existing in the prior art, and has significant application value and prospects for promotion. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the PHEMT defect detection method of the present invention; Figure 2 This is a schematic diagram of the PHEMT defect detection system of the present invention.

[0027] The attached figures are labeled as follows: 20. PHEMT's defect detection system; 201. Processor; 202. Memory. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0029] The defect detection method of PHEMT provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0030] Reference manual attached Figure 1 The diagram shows a flowchart of a PHEMT defect detection method provided by an embodiment of the present invention.

[0031] This invention provides a defect detection method for PHEMTs, which may include the following steps: S1: Obtain the surface image of PHEMT.

[0032] Among them, PHEMT is a semiconductor device based on high electron mobility technology. It is typically used in high-frequency, high-power electronic devices. By utilizing the high electron mobility characteristics of specific semiconductor materials, it achieves higher switching speeds and lower power consumption, and is widely used in wireless communication, radar systems and other fields.

[0033] It should be noted that capturing minute defects and details on the PHEMT surface through high-resolution images provides accurate basic data for subsequent defect detection and analysis.

[0034] S2: Preprocess the surface image to determine the target image.

[0035] It should be noted that normalization and flattening processes standardize the original surface image and remove background interference, thereby emphasizing the characteristics of the defect area. Normalization ensures consistency between different images and reduces the impact of illumination variations. Flattening eliminates background inhomogeneity, making the defect area more prominent.

[0036] In one possible implementation, S2 specifically includes: S201: Normalize the surface image.

[0037] S202: Flatten the normalized surface image to determine the target image.

[0038] S3: Calculate the gradient response image of the target image.

[0039] Among them, the gradient response image is an image obtained by calculating the gradient information of pixels in the image. It is usually used to detect areas with large brightness changes in the image and to identify edges or defects.

[0040] In one possible implementation, S3 specifically refers to: The gradient response image of the target image is calculated using the Sobel operator.

[0041] The Sobel operator is a commonly used edge detection operator, typically used to extract edges in the horizontal and vertical directions. The Sobel operator can calculate the gradient of an image through convolution operations.

[0042] It should be noted that calculating the gradient response map can effectively highlight the boundaries of defective regions and areas with significant changes in the image. By using the Sobel operator, areas with significant brightness changes in the image can be detected, which is crucial for discovering minute defects (such as cracks, scratches, etc.). The gradient response map can accurately capture the edge information of defects, reduce the influence of background noise, and provide a clear foundation for subsequent segmentation and defect localization.

[0043] S4: Perform iterative thresholding on the gradient response image to determine the final mask gradient image.

[0044] It should be noted that the defect region is accurately extracted through adaptive threshold segmentation. Through an iterative process, the segmentation threshold is gradually optimized and can be adaptively adjusted at different scales to cope with uneven lighting and complex backgrounds in the image. This ensures that the final mask gradient map retains only the defect-related region and removes irrelevant background noise, thus providing a more accurate region for subsequent defect localization and classification.

[0045] In one possible implementation, S4 specifically includes: S401: Perform adaptive thresholding on the gradient response image to determine the binary image.

[0046] Specifically, by using adaptive threshold segmentation, the threshold can be dynamically adjusted according to the grayscale information of local areas of the image, overcoming the limitations of fixed threshold segmentation in cases of uneven lighting and complex backgrounds. The adaptive segmentation method can more accurately identify defective areas in the image and obtain a binary image that can clearly distinguish the defective areas from the background.

[0047] In one possible implementation, S401 specifically includes: S4011: Construct the Gaussian weight matrix.

[0048] The Gaussian weight matrix is ​​as follows: in, G( x , y ) indicates the location ( x , y Gaussian distribution weights, e Represents an exponential function. s This represents the Gaussian kernel standard deviation. x and y Represents the spatial coordinates with the center of the convolution kernel as the origin.

[0049] S4012: Based on the Gaussian weight matrix, perform adaptive thresholding on the gradient response image to determine the binary image.

[0050] The binary image is specifically as follows: in, f bin ( x , y ) indicates the location ( x , y The binary image of ) f ( x , y ) indicates the location ( x , y The gradient response image of ), where λ represents the weight factor, and * represents the convolution operation. s g This represents the global standard deviation, where 1 indicates a potentially defective area and 0 indicates background.

[0051] S402: Determine the mask gradient image based on the binary image.

[0052] In one possible implementation, S402 specifically includes: S402 specifically includes: S4021: Use the binary image as a mask and invert the mask.

[0053] S4022: Based on the inverted mask, remove the strong gradient regions in the gradient response image and determine the mask gradient image.

[0054] It should be noted that by using the binary image as a mask and performing an inversion operation, the defect area in the image can be effectively highlighted. After inverting the mask, the strong gradient area in the gradient response map is removed, which can eliminate background noise or irrelevant high-contrast areas, retain the low-contrast areas related to the defect, and accurately locate the defect edge.

[0055] S403: Repeat steps S401 to S402 until the iteration termination condition is met, and determine the final mask gradient image.

[0056] The specific termination condition for the iteration is as follows: in, g i Indicates the first i The mask gradient information content of the next iteration, mean( MG i ) indicates the first i Mask gradient image in the next iteration MG i The mean, std( MG i ) indicates the first i Mask gradient image in the next iteration MG i standard deviation I sover This indicates the termination of the iteration. g i-1 Indicates the first i The amount of mask gradient information in -1 iterations. N Indicates the maximum number of iterations. dori The threshold representing the change in information content. turn This indicates that it is correct. false This indicates an error.

[0057] S5: Based on the target image, determine the Gaussian segmented image through Gaussian difference and fixed threshold segmentation.

[0058] It should be noted that by using Difference of Gaussians (DoG), detailed information at different scales can be highlighted, especially edge and texture information. By using fixed threshold segmentation to binarize the extracted Gaussian image, the defect area and background can be effectively distinguished, and important defect features in the image can be accurately detected, thus improving the accuracy and stability of defect detection.

[0059] In one possible implementation, S5 specifically includes: S501: Construct scale space and determine scale space image.

[0060] in, L ( x , y , a ) indicates the scale a The following is located in position ( x , y ) scale space image, G( x , y , a ) indicates the scale a The following is located in position ( x , yGaussian kernel.

[0061] S502: Determine the Gaussian image based on the scale-space image using Gaussian difference.

[0062] in, D ( x , y , a ) indicates the scale a The following is located in position ( x , y Gaussian image, k This represents a constant multiplied by a factor.

[0063] S503: Perform fixed threshold segmentation on the Gaussian segmentation image to determine the Gaussian segmentation image.

[0064] In one possible implementation, after S5 and before S6, the following is also included: Determine if the Gaussian segmentation image and the final mask gradient image can be merged. If yes, proceed to S6. Otherwise, return to S1.

[0065] It should be noted that by determining whether the Gaussian segmentation map and the final mask gradient map can be merged, more comprehensive information about the defect region can be extracted. By combining the two segmentation results, the advantages of Gaussian segmentation in multi-scale feature extraction and the accuracy of gradient map in edge detection can be fully utilized, thereby reducing false detections or missed detections and ensuring that the detected defect regions are more accurate and complete.

[0066] In one possible implementation, determining whether the Gaussian segmentation image and the final mask gradient image can be merged specifically includes: Calculate the spatial distance between the Gaussian segmentation image and the final mask gradient image.

[0067] The spatial distance is specifically: in, d Indicates spatial distance. p x and p y Represents the spatial coordinates of a Gaussian segmented image. q x and q y This represents the spatial coordinates of the final mask gradient image.

[0068] Determine if the spatial distance is less than a preset spatial distance. If so, the Gaussian segmentation image and the final mask gradient image can be merged. Otherwise, the Gaussian segmentation image and the final mask gradient image cannot be merged.

[0069] S6: Merge the Gaussian segmentation image and the final mask gradient image to determine the defect region.

[0070] Combining the advantages of Gaussian segmentation maps and mask gradient maps allows for more comprehensive and accurate localization of defect areas. Gaussian segmentation maps highlight multi-scale features, while mask gradient maps effectively capture edge information of the image. By merging these two types of images, background noise can be removed, false positives can be reduced, and the accuracy and integrity of defect areas can be improved.

[0071] S7: Extract defect features from the defect area.

[0072] It should be noted that by accurately extracting the features of the defect area (such as texture, shape, color, etc.), rich information can be provided for subsequent defect classification. These features can effectively distinguish different types of defects and improve the accuracy of classification.

[0073] S8: Based on the defect characteristics, perform defect classification through a deep residual convolutional network and output the defect category of PHEMT.

[0074] Among them, deep residual convolutional networks are a type of deep learning model that employs a residual learning mechanism. This mechanism can effectively train very deep convolutional neural networks (CNNs). By introducing residual blocks, this network avoids the gradient vanishing problem and enables information to be better transmitted within the network. The residual blocks directly add the input and output through skip connections, thereby simplifying the training of deep networks.

[0075] It should be noted that by automatically learning and optimizing defect features in images through Deep Residual Convolutional Networks (DFFN), different types of defects can be accurately distinguished. This not only avoids the limitations of manually designed features but also improves classification accuracy through multi-level feature extraction. By outputting the probability distribution of defect categories through a Softmax classifier, the network can achieve fast and high-precision multi-class defect classification. In addition, DFFN solves the gradient vanishing problem in deep network training through residual learning, enabling the network to handle more complex and diverse defect types, thereby improving the overall detection effect and reliability.

[0076] In one possible implementation, the deep residual convolutional network includes: an input layer, a residual layer, a feature fusion layer, a fully connected layer, a softmax classification layer, and an output layer.

[0077] The residual layer consists of convolutional units, batch normalization units, and the ReLU activation function.

[0078] S8 specifically includes: S801: Accepts defect features through the input layer.

[0079] S802: Input the defect features into the residual layer and output the enhancement features.

[0080] in, F () indicates the enhanced features of the output. A () represents the features obtained through the convolutional layer. X Indicates defect characteristics, W 1 and W 2 both represent convolution kernels, b 1 and b 2 represents the bias term, and ReLU represents the ReLU activation function.

[0081] S803: Input the enhanced features into the feature fusion layer, perform feature fusion, and output the fused features.

[0082] S804: Input the fused features into the fully connected layer and output the feature vector.

[0083] S805: Input the feature vector into the Softmax classification layer and output the probability distribution of each defect category.

[0084] in, p r Indicates the first r Predicted probability of class defects e Represents an exponential function. z r Indicates the first r Feature vectors corresponding to class defects z j Indicates the first j Feature vectors corresponding to class defects r , j =1,2,…, T , T Indicates the total number of defect categories; S806: Based on the probability distribution, output the defect category through the output layer.

[0085] in, This represents the final predicted defect category label, and arg max indicates taking the maximum value.

[0086] In one possible implementation, the loss function of the deep residual convolutional network is specifically the cross-entropy loss function.

[0087] The specific defect categories are: cracks, scratches, and stains.

[0088] In one possible implementation, it also includes: S9: Optimize deep residual convolutional networks using the Seagull optimization algorithm.

[0089] S9 specifically includes: S901: Initialize the seagull population and determine the initial position of each seagull, where the initial position represents the weights of the deep residual convolutional network.

[0090] S902: Use the loss function of the deep residual convolutional network as the fitness function to calculate the fitness value of each individual seagull.

[0091] S903: Update the position of each individual seagull based on its fitness value.

[0092] S903 specifically includes: S9031: Update the position of each individual seagull based on its fitness value: in, C s ( t ) indicates the first t The position of the individual seagulls in the next iteration. V This represents the coefficient of seagull movement behavior. P s ( t ) indicates the first t The position of the seagull in the next iteration. t Indicates the current iteration number. e Represents an exponential function. D max Indicates the maximum number of iterations. e The parameter indicates how to adjust the convergence speed. betarm (1,1) represents a random number generated from the Beta distribution.

[0093] S9032: Determine the optimal direction for the seagulls based on the updated positions of each individual seagull: in, M s ( t ) indicates the first t The optimal direction for the seagull in the next iteration. U This represents the global and local search performance of the equilibrium algorithm. P bs ( t ) indicates the first t The optimal position for the seagull in the next iteration.

[0094] S9033: Update the position of each individual seagull relative to the best seagull based on its direction. in, D s ( t ) indicates the first t The position of each individual seagull relative to the optimal seagull is updated in the next iteration.

[0095] S904: Calculate the fitness value of each individual seagull after the update: S905: When the fitness value of the updated seagull is greater than or equal to the fitness value of the current seagull, update the position of the current seagull. When the fitness value of the updated seagull is less than the fitness value of the current seagull, keep the position of the current seagull unchanged.

[0096] In this invention, the position of each individual seagull was updated in two stages.

[0097] S906: Repeat steps S901 to S905 until the maximum number of iterations is reached, and determine the optimal weights for the deep residual convolutional network.

[0098] It should be noted that the seagull optimization algorithm, by simulating the foraging and migration behavior of seagull flocks, can efficiently perform global search and local optimization in complex solution spaces, avoiding gradient vanishing and getting trapped in local optima, thus improving the training efficiency and accuracy of deep residual convolutional networks.

[0099] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least: In this embodiment of the invention, by calculating the gradient response map of the image, regions with significant intensity changes in the image can be identified. Through adaptive threshold segmentation, defect regions can be effectively separated from the background. The results are gradually refined through iterative segmentation, avoiding over-segmentation or under-segmentation. It adapts to defects of different types and scales, accurately segmenting defect regions, reducing the influence of background noise, and improving the accuracy and precision of segmentation. By using Gaussian difference and fixed threshold segmentation, the segmentation effect of defect regions is optimized, effectively reducing background interference and noise, and improving the accuracy of defect segmentation. By merging the Gaussian segmentation map and the final mask gradient map, defect regions can be determined more accurately, avoiding the problem of incomplete defect regions, improving the accuracy of defect region identification, and ensuring the integrity and accuracy of defect regions. Defect classification based on deep residual convolutional networks can effectively distinguish complex defect categories and avoid the limitations of manual classification rules.

[0100] Reference manual attached Figure 2The diagram shows a structural schematic of a PHEMT defect detection system provided in an embodiment of the present invention.

[0101] This invention provides a defect detection system 20 for PHEMT, including: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-described PHEMT defect detection method and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0102] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0103] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0104] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0105] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0106] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0107] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0108] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0109] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0110] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0111] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] This invention provides a readable storage medium comprising: storing a program or instructions on the readable storage medium, wherein when the program or instructions are executed by a processor, the program or instructions implement the steps of the above-described PHEMT defect detection method and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0113] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A defect detection method for PHEMT, characterized in that, Includes the following steps: S1: Obtain the surface image of PHEMT; S2: Preprocess the surface image to determine the target image; S3: Calculate the gradient response image of the target image; S4: Perform iterative threshold segmentation on the gradient response image to determine the final mask gradient image; S5: Based on the target image, determine the Gaussian segmented image by using Gaussian difference and fixed threshold segmentation; S6: Merge the Gaussian segmentation image and the final mask gradient image to determine the defect region; S7: Extract the defect features of the defective region; S8: Based on the defect characteristics, perform defect classification using a deep residual convolutional network and output the defect category of the PHEMT.

2. The defect detection method for PHEMT according to claim 1, characterized in that, The preprocessing in step S2 includes normalization and flattening.

3. The defect detection method for PHEMT according to claim 1, characterized in that, S3 calculates the gradient response image of the target image using the Sobel operator.

4. The defect detection method for PHEMT according to claim 1, characterized in that, S4 specifically includes: S401: Perform adaptive threshold segmentation on the gradient response image to determine a binary image; S402: Determine the mask gradient image based on the binary image; S403: Repeat steps S401 to S402 until the iteration termination condition is met, and determine the final mask gradient image.

5. The defect detection method for PHEMT according to claim 4, characterized in that, Specifically, S401 includes: S4011: Construct the Gaussian weight matrix: in, G ( x , y ) indicates the location ( x , y Gaussian distribution weights, e Represents an exponential function. σ This represents the Gaussian kernel standard deviation. x and y Represents the spatial coordinates with the center of the convolution kernel as the origin; S4012: Based on the Gaussian weight matrix, perform adaptive threshold segmentation on the gradient response image to determine the binary image; the binary image is specifically: in, f bin ( x , y ) indicates the location ( x , y The binary image of ) f ( x , y ) indicates the location ( x , y The gradient response image of ), where λ represents the weight factor, and * represents the convolution operation. σ g This represents the global standard deviation, where 1 indicates a potentially defective area and 0 indicates background.

6. The defect detection method for PHEMT according to claim 4, characterized in that, Specifically, S402 includes: S4021: Use the binary image as a mask and invert the mask; S4022: Based on the inverted mask, remove the strong gradient regions in the gradient response image to determine the mask gradient image.

7. The defect detection method for PHEMT according to claim 1, characterized in that, S5 specifically includes: S501: Construct scale space and determine scale space image: in, L ( x , y , a ) indicates the scale a The following is located in position ( x , y ) scale space image, G( x , y , a ) indicates the scale a The following is located in position ( x , y Gaussian kernel; S502: Based on the scale-space image, determine the Gaussian image through the Gaussian difference: in, D ( x , y , a ) indicates the scale a The following is located in position ( x , y Gaussian image, k Indicates a constant multiplied by a factor; S503: Perform the fixed threshold segmentation on the Gaussian image to determine the Gaussian segmented image.

8. The defect detection method for PHEMT according to claim 1, characterized in that, The following is included after S5 and before S6: Determine whether the Gaussian segmentation image and the final mask gradient image can be merged; if yes, proceed to S6; otherwise, return to S1.

9. The defect detection method for PHEMT according to claim 1, characterized in that, The deep residual convolutional network includes: an input layer, a residual layer, a feature fusion layer, a fully connected layer, a softmax classification layer, and an output layer; The residual layer includes convolutional units, batch normalization units, and ReLU activation functions; S8 specifically includes: S801: The defect features are received through the input layer; S802: Input the defect features into the residual layer and output the enhancement features: in, F () indicates the enhanced features of the output. A () represents the features obtained through the convolutional layer. X Indicates defect characteristics, W 1 and W 2 both represent convolution kernels, b 1 and b 2 represents the bias term, and ReLU represents the ReLU activation function; S803: Input the enhanced features into the feature fusion layer, perform feature fusion, and output the fused features; S804: Input the fused features into the fully connected layer and output a feature vector; S805: Input the feature vector into the Softmax classification layer and output the probability distribution of each defect category: in, p r Indicates the first r Predicted probability of class defects e Represents an exponential function. z r Indicates the first r Feature vectors corresponding to class defects z j Indicates the first j Feature vectors corresponding to class defects r , j =1,2,…, T , T Indicates the total number of defect categories; S806: Based on the probability distribution, output the defect category through the output layer: in, This represents the final predicted defect category label, and arg max indicates taking the maximum value.

10. A defect detection system for PHEMT, characterized in that, It includes a processor and a memory; the memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the PHEMT defect detection method as described in any one of claims 1 to 9.

Images