A surface quality detection method and system for semiconductor precision parts

Abstract

The present application relates to the technical field of image processing, more particularly, the present application relates to a kind of surface quality detection method and system for semiconductor precision parts, comprising: obtaining the image of part and pre-processing, the image after pre-processing is divided into complex region and simple region with periodic structure, and multiple one-dimensional feature sequences of the complex region are extracted.The present application focuses on evaluating the separation degree contribution of each data cluster from normal sample when distribution presents multimodal morphology, and focuses on the overall Fisher discriminant ratio when distribution is simple unimodal.This kind of dynamic fusion weight determination mode enables the algorithm to adaptively process defects of different complexity, effectively solving the problem that traditional methods are not sensitive to some key but small proportion defect features due to the inability to evaluate multimodal distribution information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a surface quality detection method and system for semiconductor precision parts. BACKGROUND

[0002] Semiconductor precision parts are key components of the semiconductor industry, and the surface quality of the parts directly affects the performance and reliability of chips. Such parts usually have complex periodic fine structures, such as tiny grooves or heat dissipation fins. During production, tiny defects such as scratches, burrs, and broken teeth may occur. Traditional manual visual inspection is inefficient and highly subjective, and cannot meet the detection needs of large-scale and high-precision production. Therefore, automated surface quality detection technology based on machine vision has become an inevitable trend in the industry.

[0003] In automated detection, two-dimensional image information of the part surface is usually converted into one-dimensional feature sequences for analysis. A perfect part should have a highly regular feature sequence. When the surface has defects, the regularity is destroyed, resulting in random and abrupt components in the sequence. Therefore, how to accurately quantify the complexity of these one-dimensional feature sequences and reliably determine whether there are defects based on the complexity is a problem.

[0004] To solve the above problems, permutation entropy algorithm is often used in the prior art. This algorithm can effectively measure the complexity and randomness of the sequence by analyzing the ordering mode of adjacent points in the data sequence, and is very suitable for detecting abnormalities in periodic structures. However, in practical applications, in order to improve the robustness of detection, multiple feature sequences are usually extracted and their respective permutation entropies are calculated. How to effectively integrate these multi-dimensional permutation entropy information becomes a new difficulty. The commonly used methods, such as feature selection or weighting based on Fisher discriminant ratio, assume that the permutation entropy distribution of the defect sample presents a simple unimodal pattern. However, in complex industrial scenarios, different types and degrees of defects may lead to the formation of multiple discrete data clusters, i.e. multi-modal pattern. At this time, the traditional Fisher discriminant ratio can only measure the overall separability between classes, and cannot effectively evaluate the complex multi-modal distribution information, which may lead to insensitivity to some key but small proportion of defect features, thereby affecting the accuracy of detection. SUMMARY

[0005] The application provides a surface quality detection method and system for semiconductor precision parts, aiming to solve the problem in the prior art that the traditional Fisher discriminant ratio can only measure the separability between the whole classes, cannot effectively evaluate the complex multi-peak distribution information, and may cause insensitivity to some key but small proportion defect features, thereby affecting the accuracy of detection.

[0006] In a first aspect, the application provides a surface quality detection method for semiconductor precision parts, comprising: acquiring an image of a part and preprocessing, dividing the preprocessed image into a simple region and a complex region with a periodic structure, and extracting a plurality of one-dimensional feature sequences of the complex region, and calculating the permutation entropy of each one-dimensional feature sequence; obtaining the defect possibility of the complex region by weighted sum of the permutation entropy; wherein the weight of any permutation entropy is determined according to its contribution degree, and the calculation method of the contribution degree comprises: obtaining the problem sample set and the normal sample set of the permutation entropy corresponding to the feature sequence; calculating the distribution structure complexity factor of the problem sample set, the distribution structure complexity factor being used to represent the degree of deviation of the probability density function of the problem sample set from the unimodal shape; calculating the standard Fisher discriminant ratio between the problem sample set and the normal sample set, and dynamically fusing the standard Fisher discriminant ratio with the separation degree contribution of the problem sample set to obtain the contribution degree, the separation degree contribution being used to represent the separation degree of the data cluster in the problem sample set from the normal sample set, and the degree of dynamic fusion being controlled by the distribution structure complexity factor; and judging the surface quality of the part based on the defect possibility of the complex region. By calculating the distribution structure complexity factor, the distribution shape of the defect data is intelligently judged, and according to the complexity, the dominant role of the standard Fisher discriminant ratio and the separation degree contribution in the weight calculation is dynamically adjusted, so that the contribution of various defects including clusters can be accurately evaluated, and the accuracy and robustness of detection are significantly improved.

[0007] Further, judging the surface quality of the part further comprises: calculating the defect possibility of the simple region; and if any of the defect possibility of the simple region and the defect possibility of the complex region is greater than a judgment threshold, then judging that the part is unqualified. By comprehensively evaluating the defect conditions of the complex region and the simple region, full coverage detection of the surface of the part can be realized, missing detection caused by neglecting the defects of the simple region is avoided, and the comprehensiveness and reliability of detection are further improved.

[0008] Further, the defect possibility of a simple region is calculated, including: performing binaryzation processing on the simple region by using an adaptive threshold method or a global fixed threshold method to obtain a binary image, calculating the percentage of the number of all pixel points being 1 in the binary image in all pixel points of the simple region as the defect possibility of the simple region. Compared with a complex texture analysis method, this method is more direct and effective when processing a simple structure region, can reduce the overall calculation complexity of the algorithm while ensuring the detection effect, and makes it easier to deploy and apply in an industrial scene.

[0009] Further, the permutation entropy of each one-dimensional feature sequence is calculated, including: setting an embedding dimension m and a sampling interval t, sliding a window with a length of m on the one-dimensional feature sequence, sorting the data points in the window according to the numerical value to obtain a permutation pattern of the original position index, and counting the frequency of various permutation patterns in the entire sequence, and calculating the permutation entropy according to the frequency. By setting the embedding dimension m and the sampling interval t, and sliding the window on the feature sequence to sort and count the frequency, a clear and reproducible implementation path is provided for the calculation of the core feature permutation entropy, which ensures the standardization and stability of the feature extraction process and is the basis for subsequent accurate contribution calculation and defect judgment.

[0010] Further, the distribution structure complexity factor of the problem sample set is calculated, including: using kernel density estimation on the problem sample permutation entropy set to obtain a probability density function curve; detecting local maximum points of the probability density function curve as peak values, calculating a distribution structure complexity factor based on the number of detected peak values, the standard deviation of all peak values and the mean of all peak values, the distribution structure complexity factor is positively correlated with the number of peak values, and is also positively correlated with the ratio of the standard deviation of all peak values to the mean of all peak values. By generating a probability density function curve by using kernel density estimation, and quantifying the degree of deviation from a unimodal shape based on the number of detected peak values, the standard deviation and the mean, a specific mathematical basis is provided for the step of judging the distribution complexity of the sample set. This enables the system to objectively and quantitatively evaluate the distribution characteristics of the defect data.

[0011] Further, the calculation method of the separation degree contribution includes: when the distribution structure complexity factor is greater than 1, using a peak-guided clustering algorithm to divide the problem sample permutation entropy set into multiple data clusters; calculating the local Fisher discriminant ratio between each data cluster and the normal sample permutation entropy set, multiplying each local Fisher discriminant ratio by the proportion of the sample amount of the corresponding data cluster in the total amount of problem samples, and summing to obtain the separation degree contribution.

[0012] Further, the pre-processed image is divided into a complex region and a simple region with a periodic structure, including: performing edge detection on the pre-processed image, defining the region surrounded by the two edges with the maximum distance value located at the same center as the complex region, and the remaining part as the simple region.

[0013] Further, the pre-processing of the image includes: converting the image into a gray image, using a Gaussian filter to perform denoising processing on the gray image, and performing contrast enhancement processing on the denoised gray image.

[0014] Further, the value range of the determination threshold is 0.12-0.2.

[0015] The second aspect of the present application also provides a surface quality detection system for semiconductor precision parts, comprising a processor and a memory, the memory stores a computer program, and the processor executes the computer program to realize the surface quality detection method for semiconductor precision parts described in any one of the above.

[0016] Beneficial effects: By introducing the concepts of distribution structure complexity factor and separation degree contribution. This method can first judge the distribution complexity of the problem sample set. When the distribution presents a multi-peak shape, the separation degree contribution of each data cluster to the normal sample is focused on. When the distribution is a simple single peak, the overall Fisher discriminant ratio is focused on. This dynamic fusion weight determination method makes the algorithm adaptively process different complexity defect scenarios, effectively solves the problem that the traditional method is not sensitive to some key but small proportion defect features due to the inability to evaluate multi-peak distribution information, and significantly improves the accuracy and robustness of detection. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a schematic top structure diagram of a part according to an embodiment of the present application;

[0018] Figure 2 is a flowchart schematically showing quality detection of a part according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0020] As shown in Figure 1 and Figure 2 S101: Collecting a semiconductor part image and performing pre-processing.

[0021] Specifically, a high-resolution industrial camera can be employed, and combined with an illumination system such as an on-axis light source or a dark-field light source, to clearly capture the fine structural features of the part surface. The collected raw image is usually an RGB image, and to reduce the complexity of subsequent calculations and focus on the brightness information, the image is first processed to grayscale. Subsequently, to suppress the random noise that can be introduced in the image acquisition and transmission process, a smoothing processing technique such as Gaussian filtering can be used to denoise the grayscale image. Finally, to enhance the contrast between the part structure and the background, facilitating subsequent edge detection and feature extraction, the image can be processed for contrast enhancement or histogram equalization.

[0022] S102: Region division and feature sequence extraction are performed on the image.

[0023] To distinguish different regions of the part surface, the pre-processed image needs to be regionally divided. Specifically, the inner and outer contours of the part are identified by an edge detection algorithm, wherein the edge detection algorithm can use a Canny operator. The distance between any two edges located at the same center can be sequentially counted, and the intermediate region surrounded by the two edges with the maximum distance value (the two edges with the maximum distance value are y1 and y2 in the formula) is defined as a complex region with periodic fine structures, while the remaining part is defined as a simple region, and the defects in the complex region need to be accurately distinguished in the subsequent process. Among them, the center coordinates can be obtained by using a Hough circle transformation algorithm. Figure 1

[0024] Next, the two-dimensional image information of the complex region is converted into a one-dimensional data sequence to facilitate quantitative analysis. Exemplarily, the following time sequences representing the physical arrangement features can be extracted: first, identify each groove region, establish a coordinate system with the center as the origin, calculate the mean value of the coordinate values of all pixel points in the groove region, take the mean value as the center point of the groove region, then extract the angle distribution sequence, the contour sequence and the width sequence. Angle distribution sequence: along the annular path, calculate the included angle between the center points of two adjacent grooves and the connecting line of the two points, traverse all groove regions to obtain the angle distribution sequence. Contour sequence: obtain the contour perimeter of each groove region, arrange all the perimeter values in spatial order to form the contour sequence. Width sequence: measure the radial width of each groove region from the center point of each groove region, that is, the sum of the shortest distances from the center point to the two side edges of the groove region, arrange all the width values in spatial order to form the width sequence. For a perfect part, the above sequences should present high regularity and periodicity; when there are defects such as broken teeth, burrs or wear, the original regularity is destroyed, resulting in mutations and randomness in the sequences.

[0025] S103: Calculate the permutation entropy of each feature sequence.

[0026] ​The permutation entropy is used to quantify the complexity of the one-dimensional data sequence, which is based on analyzing the ordering pattern of adjacent data points in the time series to measure its randomness. The calculation process is as follows: first, for any sequence, select the embedding dimension m and the sampling interval τ. In this embodiment, the value of the embedding dimension m is preferably 3-7, and the sampling interval τ is preferably 1, to ensure the highest sensitivity to local details of the sequence. Then, slide a window of length m on the sequence, with a sliding step of 1 element, that is, the sampling interval τ, sort the m data points in the window by value, find a permutation combination of their original position indexes, which is the permutation pattern. After sliding through the entire sequence, a plurality of permutation patterns are obtained. The frequency of each permutation pattern in the entire sequence is counted, and the permutation entropy is calculated according to the following formula: , wherein represents the permutation entropy of the i-th sequence, represents the probability of the j-th permutation pattern in the entire sequence, represents the total number of all permutation patterns.

[0027] That is, a regular sequence has fewer permutation patterns and a concentrated distribution, and the calculated permutation entropy value is lower. Conversely, a sequence that becomes random due to defects has diverse permutation patterns and a uniform distribution, and the calculated permutation entropy value will increase significantly.

[0028] For example, a sequence X = [4, 7, 2, 6, 1], and the parameters are selected: embedding dimension m = 3, sampling interval τ = 1, the first window: the window of length m = 3 frames the first 3 points of the sequence: [4, 7, 2]. Sort the 3 numbers from small to large, which is 2 < 4 < 7. The original position indexes of the 3 numbers in the first window are (starting from 0): 4 is at position 0, 7 is at position 1, and 2 is at position 2. Therefore, after sorting from small to large, the permutation pattern of the original position is (2, 0, 1). Slide the window: the window slides τ = 1 position to the right to get the second window, which frames the data [7, 2, 6]. The size relationship of the 3 numbers is 2 < 6 < 7, and the original position indexes of the 3 numbers in the second window are (1, 2, 0). Slide the window again to get the third window: now the window frames the data [2, 6, 1], and the size relationship of the 3 numbers is 1 < 2 < 6, and the original position indexes in the third window are (2, 0, 1).

[0029] S104: Calculate the defect possibility of the complex region.

[0030] After calculating the permutation entropy of each sequence of angle distribution, profile and width, the defect possibility of the complex region is obtained by weighted summation. The basis of construction is to integrate the complexity information of multiple features to evaluate the defect status more comprehensively. The calculation formula is as follows: . In the formula, represents the defect possibility of the complex region, represents the permutation entropy of the i-th sequence, represents the importance of the permutation entropy of the i-th sequence, represents the total number of sequences. The importance of the permutation entropy of the i-th sequence reflects the discriminant ability of the sequence for identifying the defect region and the normal region.

[0031] Then, the importance of the permutation entropy of each sequence needs to be calculated. The calculation method is as follows: first, two sets of image data need to be collected: one is a normal sample set composed of normal part images confirmed to have no defects, and the other is a problem sample set composed of part images containing various known defects. Repeat the above steps for all images in the two sets to obtain a normal sample permutation entropy set and a problem sample permutation entropy set. When calculating the importance of the permutation entropy of any sequence, extract the width permutation entropy set of the image from the problem sample permutation entropy set as the target problem permutation entropy set, and extract the width permutation entropy set of the image from the normal sample permutation entropy set as the target normal permutation entropy set. Use kernel density estimation method to perform non-parametric probability density estimation on the target problem permutation entropy set to obtain a smooth probability density function curve. Then, perform peak detection on the curve to identify all local maximum points, i.e. the peaks of the distribution. Based on the number and relative intensity of the peaks, calculate the distribution structure complexity factor of the target problem permutation entropy set, and the calculation formula is as follows: . In the formula, is the distribution structure complexity factor of the target problem permutation entropy set. is the number of peaks detected on the probability density function curve. is the peak height of the i-th peak, i.e. the local maximum value of the probability density function. is the average peak height of all peaks.

[0032] Based on the above formula, when the distribution is an ideal single peak, , the numerator of the second term on the right side of the formula is zero, and the value is 1. As the number of peaks increases, or the difference between the heights of each peak increases, i.e. the distribution pattern is more uneven and more complex, the value of will increase accordingly. Therefore, as a dimensionless parameter, effectively quantifies the degree to which the data distribution deviates from a simple single peak pattern. ​

[0033] Then, the separation degree contribution of the permutation entropy of the sequence is calculated, and the calculation method is as follows: the calculated When it is significantly greater than 1, it indicates that the permutation entropy set of the target problem exists a multi-peak structure. At this time, it is necessary to evaluate the separation degree of each data cluster and the normal sample distribution. First, the peak value guided clustering method (such as K-Means algorithm, taking the abscissa corresponding to each peak value as the initial centroid) is used to divide the permutation entropy set of the target problem into data subsets, and each subset represents a data cluster dominated by a specific defect type. Then, for each data cluster, the local Fisher discriminant ratio between it and the target normal permutation entropy set is calculated. Finally, the weighted sum of all local discriminant ratios is obtained to obtain the cluster separation degree contribution. The construction of the index is based on: respectively considering the contribution of each problem data cluster to the overall discriminant ability, and according to the sample size of each cluster, the weight is assigned. The calculation formula is: . In the formula, is the separation degree contribution of the permutation entropy of the sequence, is the sample number in the th problem data cluster, is the total number of problem samples, and the ratio is the weight of the cluster, and are the mean and variance of the th problem data cluster, and are the mean and variance of the target normal permutation entropy set. The separation degree contribution of the permutation entropy of the sequence is the weighted average of the separation degree of each data cluster and the normal sample. For a data cluster far away from the normal sample distribution (large) and the distribution is compact (small) , the contribution of the separation degree contribution of the permutation entropy of the sequence is greater, which makes the separation degree contribution of the permutation entropy of the sequence accurately capture the information provided by even a small number of defect sample clusters with strong discriminability.

[0034] Next, for any sequence of permutation entropy, the standard Fisher discriminant ratio of the permutation entropy of the sequence is calculated, and the calculation process is as follows: the target problem permutation entropy set and the target normal permutation entropy set of the permutation entropy of the sequence are obtained, and the mean and (variance) of the two sets are calculated. The final Fisher discriminant ratio is obtained by the following way: calculating the ratio of the square of the difference between the mean values of the two sets and the sum of the variances of the two sets. The larger the ratio is, the stronger the ability of the permutation entropy of the width sequence to distinguish normal and problem samples is, and the higher the contribution is.

[0035] Finally, the distribution structure complexity factor of the permutation entropy of each sequence obtained by the above steps is used as an adjusting factor, the standard Fisher discrimination of the permutation entropy of each sequence is dynamically fused with the separation degree contribution thereof to obtain the contribution degree of the permutation entropy of each sequence, and the calculation formula is as follows: In the formula, is the contribution degree of the permutation entropy of the sequence, is the standard Fisher discrimination ratio of the permutation entropy of the sequence, is the distribution structure complexity factor of the permutation entropy set of the target problem, is the separation degree contribution of the permutation entropy of the sequence, is the base number of the natural logarithm.

[0036] According to the above formula, when the problem sample distribution is close to unimodal is approximately equal to 1, and the exponential term tends to 1, and at this time is mainly determined by . With the increase of the distribution complexity , the exponential term rapidly decays to 0, so that the value of smoothly transitions to be dominated by .

[0037] S105: Calculate the defect possibility of the simple region.

[0038] Specifically, for the simple region, an adaptive threshold method (such as the Otsu method) can be used to determine the threshold, and a binary defect mask image is generated, and in the binary image, the percentage of all pixel points being 1 in all pixel points of the simple region is calculated as the possibility of the simple region, and the entire detection process is completed.

[0039] S106: Detect the surface quality of the part.

[0040] Specifically, a determination threshold is set, and if any of the defect possibility of the simple region and the defect possibility of the complex region is greater than the determination threshold, it is determined that the part is unqualified, otherwise, if the defect possibility of the simple region and the defect possibility of the complex region are both less than or equal to the determination threshold, it is determined that the part is qualified. In this embodiment, the experience value range of the determination threshold is set to 0.12-0.2, and preferably, the determination threshold is 0.12.

[0041] The application also provides a surface quality detection system for semiconductor precision parts. The system comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the surface quality detection method for semiconductor precision parts according to the first aspect of the application.

[0042] The system also includes other components well known to those of ordinary skill in the art such as a communications bus and communications interfaces, the arrangement and function of which are known in the art and thus will not be described in detail herein.

[0043] In this application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. For example, the computer-readable medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random-Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), and the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described herein can be implemented using computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0044] The above-described embodiments are merely illustrative of the present application and do not limit the scope of the application. It should be noted that, for those of ordinary skill in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application.

Claims

1. A method for detecting surface quality of a semiconductor precision part, characterized by, The method comprises the following steps: acquiring an image of a part and pre-processing, dividing the pre-processed image into a simple region and a complex region with a periodic structure, and extracting a plurality of one-dimensional feature sequences of the complex region, and calculating the permutation entropy of each one-dimensional feature sequence; The defect possibility of the complex region is obtained by weighted sum of the permutation entropy; wherein the weight of any permutation entropy is determined according to its contribution degree, and the calculation method of the contribution degree comprises: obtaining the problem sample set and the normal sample set of the permutation entropy corresponding to the feature sequence; calculating the distribution structure complexity factor of the problem sample set, and the distribution structure complexity factor is used to represent the degree of deviation of the probability density function of the problem sample set from the unimodal shape, and the calculation process comprises: using kernel density estimation on the problem sample permutation entropy set to obtain the probability density function curve; detecting the local maximum points of the probability density function curve as peak values, and calculating the distribution structure complexity factor based on the number of detected peak values, the standard deviation of all peak values and the mean of all peak values, the distribution structure complexity factor is positively correlated with the number of peak values, and is positively correlated with the ratio of the standard deviation of all peak values to the mean of all peak values; The standard Fisher discriminant ratio between the problem sample set and the normal sample set is calculated, and the standard Fisher discriminant ratio is dynamically fused with the separation degree contribution of the problem sample set to obtain the contribution degree, and the calculation formula is: , wherein, is the contribution degree of the permutation entropy of the sequence, is the standard Fisher discriminant ratio of the permutation entropy of the sequence, is the distribution structure complexity factor of the target problem permutation entropy set, is the base number of the natural logarithm; To contribute to the separation of the permutation entropy of the sequence, the calculation method comprises: when the distribution structure complexity factor is greater than 1, the problem sample permutation entropy set is divided into a plurality of data clusters by using the peak value guided clustering algorithm; the local Fisher discriminant ratio between each data cluster and the normal sample permutation entropy set is calculated, and the separation degree contribution is obtained by multiplying each local Fisher discriminant ratio with the proportion of the sample amount of the corresponding data cluster in the total problem sample amount and then summing up; the separation degree contribution is used to represent the separation degree of the data clusters in the problem sample set from the normal sample set, and the degree of dynamic fusion is controlled by the distribution structure complexity factor; based on the defect possibility of the complex region, the surface quality of the part is determined.

2. The surface quality inspection method for a semiconductor precision part according to claim 1, characterized by, The determination of the surface quality of the part further comprises: calculating the defect possibility of the simple region; if any of the defect possibility of the simple region and the defect possibility of the complex region is greater than a determination threshold, it is determined that the part is unqualified.

3. The surface quality inspection method for a semiconductor precision part according to claim 2, characterized by, calculating the defect possibility of the simple region comprises: using an adaptive threshold method or a global fixed threshold method to perform binaryzation processing on the simple region to obtain a binaryzation image, calculating the percentage of the number of all pixel points with a value of 1 in the simple region to the total number of pixel points in the simple region as the defect possibility of the simple region.

4. The surface quality inspection method for a semiconductor precision part according to claim 1, characterized by, calculating the permutation entropy of each one-dimensional feature sequence comprises: setting the embedding dimension m and the sampling interval t, sliding a window with a length of m on the one-dimensional feature sequence, sorting the data points in the window according to the numerical value to obtain a permutation pattern of the original position index, and counting the frequency of each permutation pattern in the entire sequence, and calculating the permutation entropy according to the frequency.

5. The surface quality inspection method for a semiconductor precision part according to claim 1, characterized by, dividing the pre-processed image into a complex region with a periodic structure and a simple region comprises: performing edge detection on the pre-processed image, and defining the region surrounded by the two edges with the maximum distance value located at the same center as the complex region, and the remaining part as the simple region.

6. The surface quality inspection method for a semiconductor precision part according to claim 1, characterized by, The pre-processing of the image comprises: The image is converted into a gray image, a Gaussian filter is used to perform denoising processing on the gray image, and contrast enhancement processing is performed on the denoised gray image.

7. The surface quality inspection method for a semiconductor precision part according to claim 2, characterized by, The determination threshold value ranges from 0.12 to 0.

2.

8. A surface quality detection system for semiconductor precision parts, comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to realize the surface quality detection method for the semiconductor precision part according to any one of claims 1-7.

Images