Wafer photoresist image acquisition device and defect detection method

Abstract

A wafer photoresist image acquisition device and defect detection method relate to the field of semiconductor integrated circuit manufacturing. The present invention addresses the issues of existing wafer photoresist defect detection methods, which still lack automated detection and exhibit low detection accuracy. The present invention comprises: designing a wafer photoresist defect image acquisition device using a high-brightness, high-uniformity, large-scale surface light source, utilizing a ring light and a surface light source to form bright and dark field illumination to acquire a wafer image; performing an improved Fourier filter on the wafer image to obtain an enhanced wafer image; binarizing the enhanced wafer image using the Otsu algorithm optimized using the moth-to-flame method to obtain a binarized wafer image; optimizing the Sobel operator using the PLIP model to obtain pixel gradient values ​​in the binarized wafer image; calculating the edge sub-gradient values ​​in the wafer image, and using an edge detection function to obtain defect outlines in the wafer image. The present invention is used to detect wafer photoresist defects.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor integrated circuit manufacturing, and in particular to a wafer photoresist image acquisition device and a defect detection method. Background Art

[0002] With the rapid development of information technology, the integrated circuit industry has continuously advanced chip manufacturing technology. The photolithography process is a crucial step in integrated circuit manufacturing, determining the fabrication of microstructures such as graphics and circuits on the chip. During the photolithography process, photoresist is used to transfer optical patterns, projecting the designed pattern onto the silicon wafer. However, photoresist is often susceptible to various factors during the manufacturing process, leading to various defects such as bubbles, particles, and cracks on the wafer surface. These defects not only affect chip production quality but can also lead to reduced product performance, increased failure rates, and even threaten the reliability of the entire electronic product. Therefore, efficient and accurate detection and identification of wafer photoresist defects has become an urgent need to improve the quality and stability of chip manufacturing.

[0003] Traditional wafer photoresist defect detection methods are mainly achieved through manual visual inspection or simple image processing methods. However, this method has low detection accuracy and cannot achieve automated detection. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem that the existing wafer photoresist defect detection method still cannot realize automated detection and has low detection accuracy, and proposes a wafer photoresist image acquisition device and defect detection method.

[0005] A wafer photoresist image acquisition device, comprising: an outer frame, a surface light source, a ring light source, an industrial camera, and a transmission mechanism;

[0006] The outer frame includes: four identical support columns, two identical fixing rods, a cover plate, and two identical connecting rods;

[0007] The support column, connecting rod and fixed rod are all provided with mounting grooves; one end of the support column is fixed to the ground, and the other end is fixedly connected to one end of the connecting rod; two support columns and a connecting rod form an outer frame cross section; the two outer frame cross sections are symmetrically arranged; the cover plate is clamped in the mounting grooves of the two connecting rods; the fixed rod is arranged between the two support columns through the mounting grooves of the support columns; the transmission mechanism is arranged in the ground area surrounded by the support columns, and the length direction of the transmission mechanism is parallel to the fixed rod; one end of the surface light source is fixed on the cover plate, and the other end is clamped in the mounting groove of the fixed rod; the surface light source is arranged toward the transmission mechanism; the annular light source is fixed on the cover plate and arranged toward the transmission mechanism; the industrial camera is clamped in the mounting groove of the fixed rod and arranged toward the transmission mechanism.

[0008] A wafer photoresist defect detection method includes the following steps:

[0009] Step 1: Use the wafer photoresist defect image acquisition device to obtain the wafer image f1(x1,y1):

[0010] f1(x1,y1)=i1(x1,y1)×r1(x1,y1) (1)

[0011] Where (x1, y1) is the pixel coordinate in the wafer image, i1(x1, y1) is the incident vector, and r1(x1, y1) is the reflected vector;

[0012] Step 2: Filter the wafer image f1(x1, y1) to obtain an enhanced wafer image;

[0013] Step 3: performing adaptive threshold binarization on the enhanced wafer image to obtain a binarized wafer image;

[0014] Step 4: Obtain edge pixel gradient values ​​in the binarized wafer image;

[0015] Step 5: Use the edge gradient value in the wafer image to obtain the defect contour in the wafer image.

[0016] Furthermore, in step 2, the wafer image f1(x1, y1) is filtered to obtain an enhanced wafer image, specifically:

[0017] Step 21: Obtain the filter transfer function H(u, v), and use the transfer function H(u, v) to separate the incident vector and the reflected vector in the wafer image, specifically:

[0018] H(u,v)F1(u,v)=H(u,v)I1(u,v)+H(u,v)R1(u,v) (2)

[0019] Among them, H(u, v) takes H h (u, v) or H l (u,v),H h (u, v) is the high-pass filter transfer function, H l (u, v) is the low-pass filter transfer function. When H(u, v) takes H h (u, v) then formula (2) is the incident component, when H(u, v) takes H l(u, v) then formula (2) is the reflection component, F1(u, v) is ln[f1(x1, y1)] after Fourier transformation, I1(u, v) is ln[i1(x1, y1)] after Fourier transformation, R1(u, v) is ln[r1(x1, y1)] after Fourier transformation, (u, v) is the pixel coordinates in the wafer image after Fourier transformation;

[0020] Step 22: Obtain filtering results of the incident vector and the reflected vector respectively;

[0021] Step 2: Utilize the filtering results of the incident vector and the reflected vector to obtain the enhanced wafer image G(x1, y1).

[0022] Furthermore, the filter transfer function H(u, v) is a high-pass filter transfer function H h (u, v) or low-pass filter transfer function H l (u, v), specifically:

[0023]

[0024]

[0025] D0=(u0,v0) (5)

[0026]

[0027] Wherein, D0 is the cutoff frequency, u0 is the horizontal coordinate of the cutoff frequency, v0 is the vertical coordinate of the cutoff frequency, k1 is the image enhancement degree control parameter, and D(u, v) is the distance from (u, v) to (u0, v0).

[0028] Furthermore, the filtering processing results of the incident vector and the reflected vector are obtained in step 22, specifically:

[0029]

[0030] h1 f1 (x1,y1)=h1 i1 (x1,y1)+h1 r1 (x1,y1) (8)

[0031] Among them, when H(u,v) takes H h (u, v), g1(x1, y1) is the result of low-pass filtering g12(x1, y1), when H(u, v) takes H l (u, v), g1(x1, y1) is the result of high-pass filtering g11(x1, y1), h1 f1(x1, y1) is the result of IFFT transformation of H(u, v)F1(u, v), h1 i1 (x1, y1) is the result of IFFT transformation of H(u, v)I1(u, v), h1 r1 (x1,y1) is the result of IFFT transformation of H(u,v)R1(u,v).

[0032] Furthermore, the enhanced wafer image G(x1, y1) is obtained by using the filtering results of the incident vector and the reflected vector in steps 2 and 3, specifically:

[0033] G(x1,y1)=g11(x1,y1)+τ×g12(x1,y1) (9)

[0034] Among them, τ is the control parameter of the image reflection vector; the value range of τ is [0.5,1].

[0035] Furthermore, in step three, the enhanced wafer image is subjected to adaptive threshold binarization to obtain a binarized wafer image, specifically:

[0036] Step 31: Prune the enhanced wafer image grayscale threshold and obtain a pruned wafer image grayscale threshold set:

[0037] Set the grayscale threshold set [t1, t2, ..., t m1 ], the maximum inter-class variance method is used to search in the gray threshold set to obtain the number of pixels n1 k2i2 ≠0 gray threshold set K={K0,K1,…,K k2}, that is, the grayscale threshold set of the pruned wafer image;

[0038] Where k2+1 is the total number of elements in the grayscale threshold set K;

[0039] Step 3.2: Use the grayscale threshold set of the pruned wafer image as moths to construct a moth population Use flames to record the best position of the target moth, and the flame set is The moth-to-fire algorithm is used to update the moth positions to obtain the optimal moth population K'1;

[0040] Among them, n2 is the population size, d is the variable dimension, is the n2th moth population, is the n2th flame, and the optimal threshold t1 is the moth with the largest weight;

[0041] The moth updates the position via:

[0042]

[0043]

[0044] Among them, D1 i is the distance between the position of the i3th moth and the position of the j'1th flame, b4 is a constant related to the spiral shape, t2 * is a random number with a value range of [-1,1], i3∈[1,n2], K1 i3 is the position of the i3th moth, i3∈[1,n2], ξ is the weight factor, is the position of the j'1th flame, ξ min is the minimum weight, the value range is [0.3,0.4], ξ max is the maximum weight, ranging from [0.5, 0.7], k3 is the weight parameter, i4 is the order of the current moths according to the fitness value, NP is the number of all moths, T2 is the maximum number of iterations, and t3' is the current number of iterations;

[0045] The number of flames is updated by:

[0046]

[0047] Among them, firenunber is the number of flames, N2 is the maximum number of flames, and round() is the rounding function;

[0048] Step 3. Based on the inter-class variance method, the optimal moth population K'1 is used to obtain the true optimal grayscale threshold, thereby obtaining a binarized image.

[0049] Furthermore, the inter-class variance of the pruned wafer images is:

[0050]

[0051]

[0052]

[0053]

[0054] Among them, ω n’1 is the percentage of pixels at each gray threshold in the pruned wafer image, i'2 is the number of gray thresholds in the pruned wafer image, n is the total number of gray thresholds in the pruned wafer image, P i'2 is the grayscale threshold t in the wafer image after pruning i'2 Probability of occurrence, μ t1is the average grayscale value of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, L'1 is the number of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, i"1 is the pixel label greater than or equal to the true optimal grayscale threshold in the pruned wafer image, P i”1 is the probability that the grayscale threshold i″1 appears in the pruned wafer image, σ b3 2 is the inter-class variance of the pruned wafer images, ω i’2 is the grayscale threshold t in the wafer image after pruning i'2 The corresponding pixel ratio, μ i’2 is the grayscale threshold t i'2 The average grayscale value of the entire pixel of the wafer image after the wafer image is binarized.

[0055] Furthermore, the step 4 of obtaining the edge pixel gradient value in the binarized wafer image is specifically as follows:

[0056] Step 4. Obtain the edge pixel gradients G0, G1 at the directions of 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° of the wafer image after binarization. 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 , specifically:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] Among them, α1, α2, α3, α4, α5, α6, α7, α8, and α9 are the pixel values ​​in the binarized wafer image, and soble is the Sobel operator;

[0067] Step 42: Use G0 and G obtained in step 41 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 Get the edge pixel gradient value G2(x1,y1) in the binary wafer image P , specifically:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] Among them, G Pa , G Pb , G Pc , G Pd is an intermediate variable.

[0074] Furthermore, the step 5 of obtaining the defect contour in the wafer image by using the edge gradient value in the wafer image is specifically as follows:

[0075] Step 51: Use the edge pixel gradient value in the wafer image to obtain the sub-pixel gradient, specifically:

[0076] When |dy2|>|dx2|:

[0077] ε=|dx2| / |dy2| (30)

[0078] d1=ε·m1+(1-ε)·m2 (31)

[0079] d2=ε·m3+(1-ε)·m4 (32)

[0080] When |dx2|>|dy2|:

[0081] ε=|dy2| / |dx2| (33)

[0082] d1=ε·m1+(1-ε)m2 (34)

[0083] d2=ε·m3+(1-ε)·m4 (35)

[0084] Where dx2 is the gradient value of the pixel with the horizontal coordinate x2, dy2 is the gradient value of the pixel with the vertical coordinate y2, ε is the weight coefficient, d1 and d2 are the gradients of the sub-pixel points, m1 and m2 are the gradients of the two pixels adjacent to d1, and m3 and m4 are the gradients of the pixels adjacent to d2.

[0085] Step 52: Compare the sub-pixel gradient with the gradient of the center point C. If C>d1 and C>d2, retain the current sub-pixel gradient value; otherwise, set the current sub-pixel gradient value to zero.

[0086] The gradient of the center point C is specifically:

[0087] C=(m1+m3) / 2+(m2+m4) / 2 (36)

[0088] Step 53: Use the Findcounter algorithm to perform contour scanning on the wafer image processed in step 52, obtain the defect contour based on the preset defect area, length, and curvature, and extract the wafer photoresist defects based on the defect contour.

[0089] The beneficial effects of the present invention are:

[0090] The present invention collects wafer images through the proposed wafer photoresist defect acquisition device. The image is filtered using an improved homomorphic filtering method, reducing wafer image noise and removing redundant information. The present invention uses an adaptive threshold binarization method to perform a binarization operation on the image, thereby improving the defect edge characteristics. The present invention uses an eight-directional Sobel gradient operator based on the PLIP model to detect image edge gradients, while optimizing it using improved non-maximum suppression, and finally uses a defect contour extraction method to obtain wafer photoresist defects. The present invention can quickly and accurately identify various defects in the photoresist, such as bubbles, particles, and cracks, during the manufacturing process, thereby realizing the automation of photoresist defect detection without manual intervention and reducing labor costs. At the same time, the present invention also improves the accuracy of wafer photoresist defect detection. In addition, the present invention has a short detection time, thereby improving the efficiency of wafer photoresist defect detection per unit time. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a device for collecting defects in wafer photoresist;

[0092] Figure 2 It is the homomorphic filtering flow chart;

[0093] Figure 3 This is a flow chart of the improved homomorphic filtering algorithm of the present invention;

[0094] Figure 4 This is the flow chart of the adaptive threshold binarization algorithm;

[0095] Figure 5 is the operator gradient template in eight directions;

[0096] Figure 6 Flowchart for detecting image edge gradients using the eight-directional Sobel gradient operator based on the PLIP model;

[0097] Figure 7 Improved non-maximum suppression method and defect contour extraction flow chart;

[0098] Figure 8 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0099] Specific implementation method 1. Figure 8 As shown, a method for detecting photoresist defects on a wafer includes the following steps:

[0100] Step 1: Use a wafer photoresist defect image acquisition device, such as Figure 1 As shown, the wafer image f1(x1,y1) is obtained:

[0101] f1(x1,y1)=11(x1,y1)×r1(x1,y1) (1)

[0102] Where (x1, y1) is the pixel coordinate in the wafer image, i1(x1, y1) is the incident vector, and r1(x1, y1) is the reflected vector;

[0103] The incident vector corresponds to low-frequency components and is reflected in areas of the image where grayscale values ​​change slowly, such as the background. Conversely, the reflected vector corresponds to high-frequency components and is reflected in areas of the image where grayscale values ​​change dramatically. These components often represent image details, such as edges. Therefore, in practical applications, this component should be enhanced.

[0104] like Figure 1 As shown, the wafer photoresist defect image acquisition device includes: an outer frame 1, a surface light source 2, a ring light source 3, an industrial camera 4, and a transmission mechanism 5;

[0105] The outer frame 1 comprises: four identical support columns, two identical fixing rods, a cover plate, and two identical connecting rods;

[0106] The support column, connecting rod and fixed rod are all provided with mounting grooves; one end of the support column is fixed to the ground, and the other end is fixedly connected to one end of the connecting rod; the two ends of the connecting rod are respectively connected to a support column; two support columns and a connecting rod constitute an outer frame cross section; the two outer frame cross sections are symmetrically arranged; the cover plate is clamped in the mounting grooves of the two connecting rods; the fixed rod is arranged between the two support columns through the mounting grooves on the support columns; the transmission mechanism 5 is a conveyor belt; the transmission mechanism 5 is arranged in the ground area surrounded by the support columns, and the length direction of the transmission mechanism 5 is parallel to the fixed rod; one end of the surface light source 2 is fixed on the cover plate, and the other end is clamped in the mounting groove of the fixed rod; the surface light source 2 is arranged toward the transmission mechanism 5; the annular light source 3 is fixed on the cover plate and arranged toward the transmission mechanism 5; the industrial camera is clamped in the mounting groove of the fixed rod and arranged toward the transmission structure 5.

[0107] Working principle: Place the wafer with defects to be identified on the transmission mechanism 5. Since the photoresist on the wafer surface is very thin, the unidirectional light beam cannot enable the camera to identify all the defects. Therefore, the present invention uses high brightness (brightness> 2000cm / m 2 ) A large-size (62cmX38cm) surface light source enables the light source to cover the entire wafer and is compatible with 6-inch, 8-inch and 12-inch wafers. Since the wafer surface has high reflective properties, the present invention adopts a side-mounted form, and the light path is a reflective light path. A high-brightness ring light is hung above the wafer. The ring light adopts a high-brightness ring cold light source, and the light beam is warm yellow light to prevent the photoresist from solidifying. The light beam illumination is greater than 200,000Lx, realizing dark field illumination. The transmission mechanism transmits the wafer with defects to be identified into the recognition range of the industrial camera, and the industrial camera obtains the wafer image.

[0108] The photoresist defect acquisition device proposed in the present invention uses a high-brightness surface light display and a ring light source to construct an environment that combines light and dark fields, and uses strip light to illuminate the high-brightness surface light display, and then uses an industrial camera to capture wafer images.

[0109] Step 2: Figure 3 As shown in FIG, the wafer image f1(x1, y1) is filtered using the improved homomorphic filtering method to obtain an enhanced wafer image, specifically:

[0110] Step 2.1: Obtain the filter transfer function and use the transfer function H(u,v) to separate the incident vector and the emitted component vector in the wafer image:

[0111] The transfer function H(u,v) is the high-pass filter transfer function H h (u,v) or low-pass filter transfer function H l (u,v);

[0112] Step 2: Get the transfer function:

[0113] First, the selection of the transfer function H(u,v) is particularly important in this process. The present invention proposes to use the deformed tanh function f2(x2) as the high-pass filter transfer function H h (u,v):

[0114] The deformed tanh function f2(x2) is as follows:

[0115]

[0116] Among them, x2 is the parameter variable;

[0117] Construct the high-pass filter transfer function H h (u,v), as follows:

[0118]

[0119] D0=(u0,v0) (4)

[0120]

[0121] Where D0 is the cutoff frequency, u0 is the abscissa of the cutoff frequency, v0 is the ordinate of the cutoff frequency, k1 is the image enhancement control parameter, (u, v) is the pixel coordinate in the wafer image after Fourier transformation, and D(u, v) is the distance from (u, v) to (u0, v0);

[0122] By function H h (u, v) processing can realize image blurring, highlight image details, and enhance high-frequency or low-frequency information of the image.

[0123] Then, using the characteristics of the simplex function, the corresponding low-pass filter transfer function H is designed. l (u, v) is as follows:

[0124]

[0125] Through the low-pass filter H l (u, v), the background information of the image can be extracted. In this way, the frequency division processing of the image is realized.

[0126] Step 212: Use the transfer function H(u, v) to separate the incident vector and the emitted component vector in the wafer image. Specifically:

[0127] First, in the frequency domain, the homomorphic filtering method can be used to adjust the brightness and contrast of the image to achieve image enhancement. Specifically, take the logarithm of both sides of formula (1), as follows:

[0128] ln[f1(x1,y1)]=ln[i1(x1,y1)]+ln[r1(x1,y1)] (7)

[0129] Then, perform Fourier transform on both sides of formula (7) to obtain the following formula:

[0130] F1(u,v)=I1(u,v)+R1(u,v) (8)

[0131] Where F1(u, v) is ln[f1(x1, y1)] after Fourier transformation, I1(u, v) is ln[i1(x1, y1)] after Fourier transformation, R1(u, v) is ln[r1(x1, y1)] after Fourier transformation, and (u, v) is the coordinate of the pixel point in the image after Fourier transformation;

[0132] Finally, the frequency domain function H(u, v) is used to multiply both sides of formula (8) to obtain the following formula:

[0133] H(u,v)F1(u,v)=H(u,v)I1(u,v)+H(u,v)R1(u,v) (9)

[0134] When H(u, v) takes H h (u, v) then formula (9) is the incident component, when H(u, v) takes H l (u, v) then formula (9) is the reflection component, so that the incident component of the image signal is separated from the reflection amount;

[0135] Step 22: Obtain the filtering results of the incident vector and the reflected vector respectively, specifically:

[0136] First, perform IFFT transformation on formula (9) and return it to the spatial domain to obtain:

[0137] h1 f1 (x1, y1) = h1 i1 (x1, y1)+h1 r1 (x1, y1) (10)

[0138] Perform exponential operations on both sides of formula (10), as follows:

[0139]

[0140] Among them, when H(u, v) takes H h (u, v), g1(x1, y1) is the result of low-pass filtering g12(x1, y1), when H(u, v) takes H l (u, v), g1(x1, y1) is the result of high-pass filtering g11(x1, y1);

[0141] Step 2 and 3: Obtain the enhanced wafer image G(x1, y1) using the filtering results of the incident vector and the reflected vector:

[0142] G(x1,y1)=g11(x1,y1)+τ×g12(x1,y1) (12)

[0143] Among them, τ is the control parameter of the image reflection vector; the value range of τ is [0.5,1].

[0144] The processing flow of this step is as follows Figure 2-3 As shown, H h ,H l The image is filtered using the transfer function of the filter, and then the processed results are inverse Fourier transformed and subjected to exponential operations respectively; the low-frequency component is then multiplied by the control parameter τ to control the grayscale value of the image. Finally, the processed g11(x1, y1) and g12(x1, y1) are linearly added to obtain the low-light image enhancement result G(x1, y1);

[0145] Step 3: Adopt the Otsu (maximum between-class variance method) method and the MFO (moth-to-flame algorithm) method to perform adaptive threshold binarization on the enhanced wafer image to obtain the binarized wafer image, specifically:

[0146] Step 31: Prune the enhanced wafer image grayscale threshold and obtain a pruned wafer image grayscale threshold set, specifically:

[0147] First, a preset optimal grayscale threshold T1 is set, and the initialized optimal grayscale threshold is used to divide the background and target of the enhanced wafer image. The inter-class variance of the enhanced wafer image is as follows:

[0148]

[0149]

[0150]

[0151]

[0152] P i2 =n1 i2 / N1 (17)

[0153] P i'1 =n1 i'1 / N1 (18)

[0154] Among them, T1 is the preset optimal grayscale threshold, L1 is the number of pixels greater than or equal to the optimal grayscale threshold, i2 is the label of the grayscale threshold, Pi2 is the gray value t i2 The probability of occurrence, i'1 is the pixel label greater than or equal to the optimal threshold, P i'1 is the grayscale threshold t i'1 Probability of occurrence, n1 i'1 is the number of pixels with grayscale threshold value i'1, n1 i2 is the grayscale threshold t i2 N1 is the total number of pixels in the enhanced wafer image, ω n1 is the pixel ratio of each gray threshold, μ n1 is the average gray value of the pixels corresponding to each gray threshold, μ T1 is the average gray value of pixels greater than the optimal gray threshold, σ b1 2 is the inter-class variance of the enhanced wafer image, m1 is the total number of image grayscale classifications, [t1, t2, …, t m1 ] is a set of grayscale thresholds, ω i2 is the grayscale threshold t i2 The corresponding pixel ratio, μ i2 is the grayscale threshold t i2 The average gray value of the entire pixel of the image after binarization;

[0155] Then, the grayscale segmentation threshold search range is pruned to narrow the search range and improve the search efficiency, and the pruned image pixel value set is obtained. Specifically, Otsu uses an exhaustive method. As the number of thresholds and grayscale ranges increase, the amount of calculation increases, so it is necessary to improve the algorithm calculation efficiency. Since the number of pixels corresponding to some grayscale values ​​in the collected image is 0, for the number of pixels in the image n1 k2 The gray value k2 is 0, then the probability of the gray value k2 appearing

[0156] Assume that in [t1,t2,…,t m1 ] Select grayscale thresholds a1 and b2 as preset thresholds, set grayscale value j, a1 <j<b2,n1 j ≠0, and n1 j+1 =0,…,n1 b2 =0, then:

[0157]

[0158]

[0159] Among them, j1 is the gray value between a1+1 and b2, j2 is the gray value between a1+1 and j, P j1 is the probability of gray value j1 appearing, P j is the probability of gray value j appearing, Pj+1 is the probability of gray value j+1 appearing, P b2 is the probability of gray value b2 appearing, ω j is the percentage of pixels with gray value j, ω b2 is the pixel ratio corresponding to the gray value b2, μ b2 is the average gray value of the entire pixel of the image after the image is binarized according to the gray threshold b2, P j2 is the probability of gray value j2 appearing, μ j It is the average gray value of the entire pixel of the image after the image is binarized according to the gray threshold j;

[0160] Therefore, the results of b2 as the threshold and j as the threshold for the inter-class variance are equivalent. Therefore, the search efficiency can be improved by pruning the threshold search range to narrow the search range. The specific operation is: find the number of pixels n1 k2i2 ≠0 gray threshold set K={K0,K1,…,K k2}, the final optimal threshold t1 is searched in K. The inter-class variance σ of the pruned wafer image b3 2 The calculation formula is as follows:

[0161]

[0162]

[0163]

[0164]

[0165] Among them, ω n’1 is the percentage of pixels at each gray threshold in the pruned wafer image, i'2 is the number of gray thresholds in the pruned wafer image, n is the total number of gray thresholds in the pruned wafer image, P i'2 is the grayscale threshold t in the wafer image after pruning i'2 Probability of occurrence, μ t1 is the average grayscale value of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, L'1 is the number of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, i"1 is the pixel label greater than or equal to the true optimal grayscale threshold in the pruned wafer image, P i”1 is the probability that the grayscale threshold i″1 appears in the pruned wafer image, σ b3 2 is the inter-class variance of the pruned wafer images, ω i’2 is the grayscale threshold t in the wafer image after pruning i'2 The corresponding pixel ratio, μ i’2 is the grayscale threshold t i'2The average grayscale value of all pixels in the wafer image after the wafer image is binarized, and k2+1 is the total number of elements in the grayscale threshold set K;

[0166] Step 3.2: Take the grayscale threshold set K as the candidate solution, and take the candidate solution as the moth. The moth population is represented as Use flames to record the best position of the target moth, and the flame set is The moth-to-fire algorithm is used to update the moth positions to obtain the optimal moth population K'1:

[0167] Among them, d is the variable dimension, n2 is the number of populations;

[0168] The formula for updating the moth's position is:

[0169] D1 i3 =|F2 j'1 -K1 i3 | (25)

[0170] K1=D1 i3 ×e b4t2* ×cos(2πt2*)+F2 j'1 (26)

[0171] Where D1 i is the distance between the position of the i3th moth and the position of the j'1th flame, b4 is a constant related to the spiral shape, usually 1, t2* is a random number with a value range of [-1,1], K1 i3 is the position of the i3th moth, i3∈[1,n2];

[0172] In order to avoid falling into local optimum, an adaptive flame number mechanism is used, and the flame number update formula is:

[0173]

[0174] Where firenunber is the number of flames, t3' is the current iteration number, N2 is the maximum number of flames, T2 is the maximum iteration number, and round is the rounding function;

[0175] In this invention, a weight factor ξ is introduced to improve the slow convergence speed of the MFO algorithm in the later stage. When the inertia weight is relatively large, the global search capability of the algorithm is improved; when the inertia weight is relatively small, the local search capability of the algorithm is improved. The calculation formula of the weight factor is shown in formula (25):

[0176]

[0177] Where ξ minis the minimum weight, the value range is [0.3,0.4], ξ max is the maximum weight, ranging from [0.5, 0.7], k3 is the weight parameter, generally 0.1, i4 is the order of the current moths according to their fitness values, and NP is the number of all moths.

[0178] The update formula of the moth position with the weight factor ζ is:

[0179] K1 i3 =ξ×D1 i3 ×e b4t2* ×cos(2πt2 * )+F2j'1 (29)

[0180] in, is the position of the j'1th flame;

[0181] By passing the K set through the improved MFO algorithm, the optimal moth population, that is, the set K'1 where the optimal threshold t1 is located, is obtained;

[0182] Step 3. Input the set K'1 of the optimal threshold t1 obtained in step 3.2 into the Otsu method (maximum inter-class variance method) to obtain the true optimal grayscale threshold t1 and obtain the binary wafer image, as shown in the following example: Figure 4 shown.

[0183] Step 4: The eight-directional Sobel gradient operator based on the PLIP model is used to obtain the edge pixel gradient value in the binary wafer image, such as Figure 6 As shown, specifically:

[0184] This design uses an eight-directional Sobel operator combined with a parameterized logarithmic image processing model (PLIP) to replace traditional linear operations. This allows for more accurate calculation of image gradients while avoiding the complexity and inaccuracy of the traditional Canny algorithm. This algorithm utilizes an eight-directional Sobel operator template for calculation, thereby more comprehensively considering image gradient information and improving algorithm accuracy.

[0185] To address the problems of poor performance and weak anti-interference capabilities often encountered by traditional edge detection algorithms when capturing high-precision images under both bright and dark field illumination, we have improved the PLIP model by adding a set of parameters. Based on the PLIP model, we derive a new gradient operator that replaces the gradient calculation of the traditional Canny operator edge detection. This model replaces linear operations with nonlinear operations, better reflecting the nonlinear characteristics of the image and providing users with more flexible parameter selection to achieve better image processing performance. This also reduces the algorithm's complexity and further improves its robustness. The PLIP model function is:

[0186] t3 # =μ3-f3 (30)

[0187]

[0188]

[0189]

[0190] Its basic homomorphic function is:

[0191]

[0192]

[0193] Where f3 is the binarized image, t3 # , t31, t32 are grayscale functions, μ3, k4, λ are arbitrary linear functions, β is the PLIP parameter, which is a scalar. is PLIP addition, -- is PLIP subtraction, is PLIP multiplication, is a homomorphic function.

[0194] The PLIP model can optimize image processing results by adjusting its parameters. Its highly adaptable parameters have broadened its application in image processing fields such as image enhancement and edge detection. In-depth research and analysis have shown that the optimal parameter selection method is typically β = 2, μ3 = k4 = λ = 1026, enabling the PLIP model to better adapt to various image processing requirements while improving both efficiency and accuracy.

[0195] When using the traditional Sobel operator template to calculate the gradient of an image, the edge information in the diagonal direction of the image will be ignored, resulting in missed detection and loss of the image edge. To address this problem, the present invention adopts the detection operator in eight directions, namely 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, specifically:

[0196] Step 4. Obtain the edge pixel gradients G0, G1 at the directions of 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° of the wafer image after binarization. 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 , specifically:

[0197]

[0198]

[0199]

[0200]

[0201]

[0202]

[0203]

[0204]

[0205]

[0206] Among them, α1, α2, α3, α4, α5, α6, α7, α8, and α9 are the pixel values ​​in the binary wafer image, and the gradient operator template in each direction is as follows: Figure 5 As shown;

[0207] Step 42: Use G0 and G obtained in step 41 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 Get the edge pixel gradient value G2(x1, y1) in the binary wafer image P , specifically:

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214] Among them, θ(x1, y1) is the gradient direction angle, G Pa , G Pb , G Pc , G Pd is an intermediate variable.

[0215] Step 5: Figure 7 As shown in FIG, based on the improved non-maximum suppression method, the edge gradient value in the wafer image is used to obtain the defect contour in the wafer image, specifically:

[0216] The traditional Canny operator only processes gradients in the 0°, 45°, 90°, and 135° directions when performing non-maximum suppression. However, in real-world images, the edge gradients do not necessarily lie along the specified directions. Therefore, the traditional non-maximum suppression method is insufficient to accurately extract edge maxima. This invention improves the accuracy of non-maximum suppression and reduces the risk of edge point misoperation by using linear interpolation to calculate the gradient values ​​of sub-pixel points in the image.

[0217] Step 51: Use the edge pixel gradient value in the wafer image to obtain the sub-pixel gradient, specifically:

[0218] When |dy2|>|dx2|:

[0219] ε=|dx2| / |dy2| (51)

[0220] d1=ε·m1+(1-ε)·m2 (52)

[0221] d2=ε·m3+(1-ε)·m4 (53)

[0222] When |dx2|>|dy2|:

[0223] ε=|dy2| / |dx2| (54)

[0224] d1=ε·m1+(1-ε)·m2 (55)

[0225] d2=ε·m3+(1-ε)·m4 (56)

[0226] Where dx2 is the gradient value of the pixel with the horizontal coordinate x2, dy2 is the gradient value of the pixel with the vertical coordinate y2, ε is the weight coefficient, d1 and d2 are the gradients of the sub-pixel points, m1 and m2 are the gradients of the two pixels adjacent to d1, and m3 and m4 are the gradients of the pixels adjacent to d2.

[0227] Step 52: Compare the sub-pixel gradient with the gradient of the center point C. If C>d1 and C>d2, it means that the center point of the current pixel is the maximum value in the neighborhood, and its gradient value is retained; otherwise, the gradient value of the current pixel is set to zero.

[0228] C=(m1+m3) / 2+(m2+m4) / 2 (57)

[0229] Step 53: Call the Findcounter algorithm in the open cv library to perform contour scanning on the image after gradient value processing, save the contour data into an array, traverse the array and use morphological features such as area, length, and curvature to filter out the desired defect contour, and then the wafer photoresist defects can be extracted.

[0230] Example:

[0231] In order to verify the beneficial effects of the present invention, the defect extraction results of the present invention are verified. When processing the identification problem, the relevant indicators can be calculated in the following way.

[0232] Table 1 Test results statistics

[0233] Confusion Matrix Number of defects identified The actual number of defects TP The actual number of defects is not FP Number of missed inspections LP

[0234] The accuracy is:

[0235]

[0236] The precision is:

[0237]

[0238] The recall rate is:

[0239]

[0240] F1 value (F1-score): It is the harmonic mean of recall and precision.

[0241] The F1 value is:

[0242]

[0243] In the present invention, data of 500 wafer photoresists were collected. After manual review, a total of 348 defects were found. A total of 342 defects were detected using the method provided by the present invention, and the obtained TP was 339, FP was 3, and LP was 9. The calculated scheme accuracy reached 96.6%, and the system F1 reached 98.3%.

Claims

1. A method for detecting photoresist defects on a wafer, characterized in that The specific process of the method is: Step 1: Use the wafer photoresist defect image acquisition device to obtain the wafer image f1(x1,y1): f1(x1,y1)=i1(x1,y1)×r1(x1,y1) (1) Where (x1, y1) is the pixel coordinate in the wafer image, i1(x1, y1) is the incident vector, and r1(x1, y1) is the reflected vector; Step 2: Filter the wafer image f1(x1,y1) to obtain an enhanced wafer image; Step 3: Adaptively threshold binarize the enhanced wafer image to obtain a binarized wafer image, specifically: Step 31: Prune the enhanced wafer image grayscale threshold and obtain a pruned wafer image grayscale threshold set: Set the grayscale threshold set [t1, t2, ..., t m1 ], the maximum inter-class variance method is used to search in the gray threshold set to obtain the number of pixels n1 k2i2 ≠0 gray threshold set K={K0,K1,…,K k2 }, that is, the grayscale threshold set of the pruned wafer image; Where k2+1 is the total number of elements in the grayscale threshold set K; Step 3.2: Use the grayscale threshold set of the pruned wafer image as moths to construct a moth population Use flames to record the best position of the target moth, and the flame set is The moth-to-fire algorithm is used to update the moth positions to obtain the optimal moth population K'1; Among them, n2 is the population size, d is the variable dimension, is the n2th moth population, is the n2th flame, and the optimal threshold t1 is the moth with the largest weight; The moth updates the position via: Among them, D1 i is the distance between the position of the i3th moth and the position of the j'1th flame, b4 is a constant related to the spiral shape, t2 * is a random number with a value range of [-1,1], i3∈[1,n2], K1 i3 is the position of the i3th moth, i3∈[1,n2], ξ is the weight factor, is the position of the j'1th flame, ξ min is the minimum weight, the value range is [0.3,0.4], ξ max is the maximum weight, ranging from [0.5, 0.7], k3 is the weight parameter, i4 is the order of the current moths according to the fitness value, NP is the number of all moths, T2 is the maximum number of iterations, and t3' is the current number of iterations; The number of flames is updated by: Among them, firenunber is the number of flames, N2 is the maximum number of flames, and round() is the rounding function; Step 3. Based on the inter-class variance method, the optimal moth population K'1 is used to obtain the true optimal grayscale threshold, thereby obtaining a binarized image; Step 4: Obtain edge pixel gradient values ​​in the binarized wafer image using an eight-directional Sobel gradient operator based on the parameterized logarithmic image processing model PLIP. Step 5: Based on the improved non-maximum suppression method, the edge gradient value in the wafer image is used to obtain the defect contour in the wafer image.

2. The method for detecting photoresist defects on a wafer according to claim 1, wherein: The wafer image f1(x1, y1) in step 2 is filtered to obtain an enhanced wafer image, specifically: Step 21: Obtain the filter transfer function H(u,v), and use the transfer function H(u,v) to separate the incident vector and the reflected vector in the wafer image. Specifically: H(u,v)F1(u,v)=H(u,v)I1(u,v)+H(u,v)R1(u,v) (2) Among them, H(u,v) takes H h (u,v) or H l (u,v),H h (u,v) is the high-pass filter transfer function, H l (u,v) is the low-pass filter transfer function, when H(u,v) takes H h (u,v) then formula (2) is the incident component, when H(u,v) takes H l (u, v) then formula (2) is the reflection component, F1(u, v) is ln[f1(x1, y1)] after Fourier transformation, I1(u, v) is ln[i1(x1, y1)] after Fourier transformation, R1(u, v) is ln[r1(x1, y1)] after Fourier transformation, (u, v) is the pixel coordinates in the wafer image after Fourier transformation; Step 22: Obtain filtering results of the incident vector and the reflected vector respectively; Step 2: Use the filtering results of the incident vector and the reflected vector to obtain the enhanced wafer image G(x1, y1).

3. The method for detecting photoresist defects on a wafer according to claim 2, wherein: The filter transfer function H(u,v) is the high-pass filter transfer function H h (u,v) or low-pass filter transfer function H l (u,v), specifically: D0=(u0,v0) (5) Among them, D0 is the cutoff frequency, u0 is the horizontal coordinate of the cutoff frequency, v0 is the vertical coordinate of the cutoff frequency, k1 is the image enhancement degree control parameter, and D(u,v) is the distance from (u,v) to (u0,v0).

4. The method for detecting photoresist defects on a wafer according to claim 3, wherein: The filtering results of the incident vector and the reflected vector are obtained in step 22, specifically: h1 f1 (x1,y1)=h1 i1 (x1,y1)+h1 r1 (x1,y1) (8) Among them, when H(u,v) takes H h (u, v), g1(x1, y1) is the result of low-pass filtering g12(x1, y1), when H(u, v) takes H l (u, v), g1(x1, y1) is the result of high-pass filtering g11(x1, y1), h1 f1 (x1, y1) is the result of IFFT transformation of H(u, v)F1(u, v), h1 i1 (x1, y1) is the result of IFFT transformation of H(u, v)I1(u, v), h1 r1 (x1,y1) is the result of IFFT transformation of H(y,v)R1(u,v).

5. The method for detecting photoresist defects on a wafer according to claim 4, wherein: The enhanced wafer image G(x1, y1) is obtained by using the filtering results of the incident vector and the reflected vector in steps 2 and 3, specifically: G(x1,y1)=g11(x1,y1)+τ×g12(x1,y1) (9) Among them, τ is the control parameter of the image reflection vector; the value range of τ is [0.5,1].

6. The method for detecting photoresist defects on a wafer according to claim 5, wherein: The inter-class variance of the pruned wafer images is: Among them, ω n’1 is the percentage of pixels at each gray threshold in the pruned wafer image, i'2 is the number of gray thresholds in the pruned wafer image, n is the total number of gray thresholds in the pruned wafer image, P i'2 is the grayscale threshold t in the wafer image after pruning i'2 Probability of occurrence, μ t1 is the average grayscale value of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, L'1 is the number of pixels greater than or equal to the optimal grayscale threshold in the pruned wafer image, i"1 is the pixel label greater than or equal to the true optimal grayscale threshold in the pruned wafer image, P i”1 is the probability that the grayscale threshold i″1 appears in the pruned wafer image, σ b3 2 is the inter-class variance of the pruned wafer images, ω i’2 is the grayscale threshold t in the wafer image after pruning i'2 The corresponding pixel ratio, μ i’2 is the grayscale threshold t i'2 The average grayscale value of the entire pixel of the wafer image after the wafer image is binarized.

7. The method for detecting photoresist defects on a wafer according to claim 6, wherein: The step 4 of obtaining the edge pixel gradient value in the binarized wafer image is specifically as follows: Step 4. Obtain the edge pixel gradients G0, G1 at the directions of 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° of the wafer image after binarization. 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 , specifically: Among them, α1, α2, α3, α4, α5, α6, α7, α8, and α9 are the pixel values ​​in the binarized wafer image, and soble is the Sobel operator; Step 42: Use G0 and G obtained in step 41 45 , G 90 , G 135 , G 180 , G 225 , G 270 , G 315 Get the edge pixel gradient value G2(x1,y1) in the binary wafer image P , specifically: Among them, G Pa , G Pb , G Pc , G Pd is an intermediate variable.

8. The method for detecting photoresist defects on a wafer according to claim 7, wherein: The step 5 of obtaining the defect contour in the wafer image by using the edge gradient value in the wafer image is specifically as follows: Step 51: Use the edge pixel gradient value in the wafer image to obtain the sub-pixel gradient, specifically: When |dy2|>|dx2|: ε=|dx2| / |dy2| (30) d1=ε·m1+(1-ε)·m2 (31) d2=ε·m3+(1-ε)·m4 (32) When |dx2|>|dy2|: ε=|dy2| / |dx2| (33) d1=ε·m1+(1-ε)·m2 (34) d2=ε·m3+(1-ε)·m4 (35) Where dx2 is the gradient value of the pixel with the horizontal coordinate x2, dy2 is the gradient value of the pixel with the vertical coordinate y2, ε is the weight coefficient, d1 and d2 are the gradients of the sub-pixel points, m1 and m2 are the gradients of the two pixels adjacent to d1, and m3 and m4 are the gradients of the pixels adjacent to d2. Step 52: Compare the sub-pixel gradient with the gradient of the center point C. If C>d1 and C>d2, retain the current sub-pixel gradient value; otherwise, set the current sub-pixel gradient value to zero. The gradient of the center point C is specifically: C=(m1+m3) / 2+(m2+m4) / 2 (36) Step 53: Use the Findcounter algorithm to perform contour scanning on the wafer image processed in step 52, obtain the defect contour based on the preset defect area, length, and curvature, and extract the wafer photoresist defects based on the defect contour.

Images