A semiconductor chip appearance defect detection method, device, system and medium

Abstract

This invention relates to the field of semiconductor chip inspection technology, and discloses a method, apparatus, system, and medium for detecting appearance defects in semiconductor chips. The method first converts an initial chip image into a corresponding spectrum diagram, fully considering the characteristics of frosted defects being non-high-frequency and edgeless. Furthermore, a first grid cell, determined through two judgment conditions, can be used to represent the coverage area of ​​the frosted defect in the spectrum diagram. Then, based on all the first grid cells, a first frequency band is obtained that can represent the predicted frequency range of frosted defects in the initial image. A second frequency band, corresponding to a low-frequency range, is used to represent the overall brightness. Simultaneously, the second frequency band includes low-frequency information corresponding to smooth interfaces. Finally, by comparing all the energies within the two frequency bands, the energy ratio can accurately represent the intensity of frosted defects, independent of illumination conditions, while fully considering the characteristics of frosted defects, thereby improving the accuracy of semiconductor chip appearance defect detection.

Description

Technical Field

[0001] This invention relates to the field of semiconductor chip inspection technology, and in particular to a method, apparatus, system, and medium for detecting appearance defects in semiconductor chips. Background Technology

[0002] In semiconductor chip defect detection, frosted defects are a relatively unique type of surface anomaly. Unlike scratches, foreign objects, or breaks, which are localized and significant defects, frosted defects are caused by microscopic roughening of the surface. These defects typically manifest as overall or regional changes in the texture details of the chip surface. Their main characteristic is not a clear edge contour, but rather an alteration in the distribution of high-frequency textures and the pattern of grayscale variations. Because they lack a distinct outline, conventional defect detection methods struggle to identify frosted defects.

[0003] During actual data acquisition, the brightness, contrast, exposure intensity, and background reflection conditions of images from different chips fluctuate to some extent. The grayscale changes of frosted defects in the spatial domain are often very weak. When there is uneven lighting, surface reflection, or local shadows, relying solely on grayscale distribution for detection is easily interfered with, resulting in inconsistent appearances of the same type of frosted defect in different images and inaccurate detection results. Summary of the Invention

[0004] This invention provides a method, apparatus, system, and storage medium for detecting appearance defects in semiconductor chips, which improves the accuracy of appearance defect detection in semiconductor chips.

[0005] The first aspect of this invention discloses a method for detecting surface defects in semiconductor chips, the method comprising: Obtain an initial chip image; based on the initial chip image, obtain a spectrum diagram, which is used to represent the image frequency and energy distribution; The spectrum is divided into grids to obtain multiple grid cells; Candidate grid cells are obtained through screening, wherein the candidate grid cells are grid cells whose energy is greater than a preset first threshold corresponding to adjacent grid cells; The first grid cell is obtained by filtering, wherein the first grid cell is a candidate grid cell whose frequency is greater than a preset second threshold and less than a preset third threshold; A first frequency band is obtained based on the first grid cell; the first frequency band is used to represent the predicted frequency range of frosted defects in the initial chip image; a second frequency band is obtained based on the spectrum and the second threshold; the second frequency band is used to represent the predicted frequency range of the overall brightness information of the initial chip image. The energy ratio is obtained by comparing all the energies corresponding to the first frequency band with all the energies corresponding to the second frequency band. The energy ratio is used to represent the degree of frosting defects in the initial chip image.

[0006] As an optional implementation, in the first aspect of the present invention, obtaining the first frequency band based on the first grid cell further includes: Based on the first grid cell, an initial frequency band is obtained, which is used to represent the frequency range corresponding to the first grid cell; The initial frequency band is subjected to an adjacency update operation: an adjacency frequency band is constructed; based on the adjacency frequency band, the initial frequency band is updated to obtain the updated initial frequency band and a first adjustment parameter, wherein the first adjustment parameter is used to measure the change in the dispersion of the initial frequency band before and after the update, and the dispersion is used to evaluate the degree of dispersion of the energy distribution corresponding to the frequency band; When it is determined that the first adjustment parameter is less than the preset fourth threshold, the updated initial frequency band is determined to be the first frequency band.

[0007] As an optional implementation, in the first aspect of the present invention, the method further includes: Based on the updated initial frequency band and the adjacent frequency band, an adjacency coordination parameter is obtained. The adjacency coordination parameter is used to measure the change in the dispersion between the adjacent frequency band and the updated initial frequency band. A second adjustment parameter is obtained based on the adjacency coordination parameter and the first adjustment parameter. The second adjustment parameter is used to measure the synchronization between the adjacency coordination parameter and the first adjustment parameter. When it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, multiple rounds of the adjacency update operation are performed on the updated initial frequency band to determine the first frequency band based on the update results.

[0008] As an optional implementation, in the first aspect of the present invention, the method further includes: Construct a secondary adjacent frequency band, wherein the secondary adjacent frequency band is a frequency band that is adjacent to the adjacent frequency band in a direction other than the initial frequency band; Based on the secondary adjacent frequency band, a secondary coordination parameter is obtained. The secondary coordination parameter is used to measure the change in the dispersion between the adjacent frequency band and the secondary adjacent frequency band. The secondary coordination parameter is used to: when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, control the upper limit of the adjacency update round to be lower than the preset fifth threshold.

[0009] As an optional implementation, in the first aspect of the present invention, the method further includes: The spectrum is inversely transformed to obtain an inverse graph, which is used to represent the corresponding positional relationship of image frequencies on the initial chip image; The inverter diagram is divided into grids to select a second grid cell, wherein the second grid cell is the grid cell corresponding to the first frequency band in the inverter diagram; Based on the second grid cell, a confidence parameter is obtained, which is used to measure the confidence that the second grid cell contains a frosted defect; based on the confidence parameter, all energies corresponding to the first frequency band are updated to obtain an updated energy ratio, which is used to represent the degree of frosted defect in the initial chip image.

[0010] As an optional implementation, in the first aspect of the present invention, the method further includes: Obtain a sample spectrum set, which includes multiple sample spectrum images, wherein the sample spectrum images are spectrum images corresponding to semiconductor chip images with frosted defects; Based on the trained detection model, the spectrogram is analyzed to obtain the third frequency band; wherein, the trained detection model is obtained by training based on the sample spectrogram set. Based on the first frequency band and the third frequency band, a consistent coordination parameter is obtained. The consistent coordination parameter is used to measure the consistency between the first frequency band and the third frequency band. The consistent coordination parameter is used to control the upper limit of the adjacent update round to be lower than the fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions.

[0011] As an optional implementation, in the first aspect of the present invention, the method further includes: Based on the current adjacency update round, the fourth threshold is updated to narrow the judgment range of the first adjustment parameter; Based on the current adjacency update round, the adjacent frequency band is updated to narrow the frequency range corresponding to the adjacent frequency band.

[0012] A second aspect of the present invention discloses a semiconductor chip appearance defect detection device, the device comprising: An image acquisition module is used to acquire an initial chip image; based on the initial chip image, a spectrum diagram is obtained, which is used to represent the image frequency and energy distribution; A grid filtering module is used to divide the spectrum into grids to obtain multiple grid cells; filter to obtain candidate grid cells, wherein the candidate grid cells are grid cells whose energy corresponding to adjacent grid cells is greater than a preset first threshold; filter to obtain a first grid cell, wherein the first grid cell is a candidate grid cell whose corresponding frequency is greater than a preset second threshold and less than a preset third threshold. A frequency band confirmation module is used to obtain a first frequency band based on the first grid cell; the first frequency band is used to represent the predicted frequency range of frosted defects in the initial chip image; and to obtain a second frequency band based on the spectrum and the second threshold, the second frequency band being used to represent the predicted frequency range of the overall brightness information of the initial chip image. An energy ratio module is used to calculate the ratio of all energies corresponding to the first frequency band to all energies corresponding to the second frequency band, thereby obtaining an energy ratio. The energy ratio is used to represent the degree of frosting defects in the initial chip image.

[0013] As an optional implementation, in a second aspect of the present invention, the frequency band confirmation module obtains the specific operation mode of the first frequency band based on the first grid cell, including: Based on the first grid cell, an initial frequency band is obtained, which is used to represent the frequency range corresponding to the first grid cell; The initial frequency band is subjected to an adjacency update operation: an adjacency frequency band is constructed; based on the adjacency frequency band, the initial frequency band is updated to obtain the updated initial frequency band and a first adjustment parameter, wherein the first adjustment parameter is used to measure the change in the dispersion of the initial frequency band before and after the update, and the dispersion is used to evaluate the degree of dispersion of the energy distribution corresponding to the frequency band; When it is determined that the first adjustment parameter is less than the preset fourth threshold, the updated initial frequency band is determined to be the first frequency band.

[0014] As an optional implementation, in a second aspect of the invention, the apparatus further includes: The synchronization adjustment module is used to obtain adjacency coordination parameters based on the updated initial frequency band and the adjacent frequency band, wherein the adjacency coordination parameters are used to measure the change in dispersion between the adjacent frequency band and the updated initial frequency band; and to obtain a second adjustment parameter based on the adjacency coordination parameters and the first adjustment parameter, wherein the second adjustment parameter is used to measure the synchronization between the adjacency coordination parameters and the first adjustment parameter; and when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, to perform multiple rounds of adjacency update operations on the updated initial frequency band to determine the first frequency band based on the update results.

[0015] As an optional implementation, in a second aspect of the invention, the apparatus further includes: The secondary coordination module is used to construct secondary adjacent frequency bands, which are frequency bands that are adjacent to the adjacent frequency bands in directions other than the initial frequency band. Based on the secondary adjacent frequency bands, secondary coordination parameters are obtained. These secondary coordination parameters are used to measure the change in dispersion between the adjacent frequency bands and the secondary adjacent frequency bands. Furthermore, the secondary coordination parameters are also used to: when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, control the upper limit of the adjacency update rounds to be lower than a preset fifth threshold.

[0016] As an optional implementation, in a second aspect of the invention, the apparatus further includes: An inverter detection module is used to perform an inverse transformation on the spectrum to obtain an inverter map, which represents the corresponding positional relationship of image frequencies on the initial chip image; to divide the inverter map into grids to filter out second grid cells, wherein the second grid cell is the grid cell corresponding to the first frequency band in the inverter map; to obtain a confidence parameter based on the second grid cell, which measures the confidence that the second grid cell contains a frosting defect; and to update all energies corresponding to the first frequency band based on the confidence parameter to obtain an updated energy ratio, which represents the degree of frosting defect in the initial chip image.

[0017] As an optional implementation, in a second aspect of the invention, the apparatus further includes: The sample detection module acquires a sample spectrum set, which includes multiple sample spectrum images, each corresponding to a semiconductor chip image with a frosted defect. Based on a trained detection model, the module analyzes the spectrum images to obtain a third frequency band. The trained detection model is obtained by training the sample spectrum set. A consistency coordination parameter is obtained based on the first and third frequency bands. This parameter measures the consistency between the first and third frequency bands and is used to control the upper limit of the adjacent update rounds to be lower than the fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions.

[0018] As an optional implementation, in a second aspect of the invention, the apparatus further includes: The range reduction module is used to update the fourth threshold based on the current adjacency update round to narrow the judgment range of the first adjustment parameter; and to update the adjacent frequency band based on the current adjacency update round to narrow the frequency range corresponding to the adjacent frequency band.

[0019] A third aspect of this invention discloses a semiconductor chip appearance defect detection system, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the semiconductor chip appearance defect detection method according to any of the first aspects of the present invention.

[0020] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked by a processor, are used to execute some or all of the steps in the semiconductor chip appearance defect detection method described in any of the first aspects of the present invention.

[0021] Compared with the prior art, the present invention has the following beneficial effects: In this invention, firstly, the initial chip image is converted into a corresponding spectrum diagram of the chip, fully considering the characteristics of frosted defects being non-high-frequency and edgeless. Furthermore, a first grid cell, determined through two judgment conditions, can be used to represent the coverage area of ​​the frosted defect in the spectrum diagram. Then, based on all the first grid cells, a first frequency band is obtained that can represent the predicted frequency range of the frosted defect in the initial image. Since the second frequency band corresponds to the low-frequency range and is used to represent overall brightness, and also includes low-frequency information corresponding to smooth interfaces, finally, by comparing all the energies within the two frequency bands, the energy ratio can accurately represent the intensity of the frosted defect, independent of illumination conditions, while fully considering the characteristics of the frosted defect, thereby improving the accuracy of semiconductor chip appearance defect detection. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic flowchart of a semiconductor chip appearance defect detection method disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a semiconductor chip appearance defect detection device disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a semiconductor chip appearance defect detection system disclosed in an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] Example 1 Please see Figure 1 , Figure 1 This is a schematic flowchart of a semiconductor chip appearance defect detection method disclosed in an embodiment of the present invention. Wherein, Figure 1 The described method for detecting surface defects in semiconductor chips can be applied to semiconductor chip surface defect detection devices. For example... Figure 1 As shown, the chip surface defect detection method may include the following operations: Step 101: Obtain the initial chip image.

[0028] In this embodiment of the invention, the initial chip image can be an image containing a semiconductor chip. This image may include a complete semiconductor chip, a portion of a semiconductor chip, or may only sample the location of the frosted defect; this embodiment of the invention does not limit this. Preferably, the initial chip image can be a dark-field image containing a semiconductor chip, which can clearly highlight extremely small surface defects such as frosted defects from a clean, highly reflective background.

[0029] It should be noted that the spatial resolution of the camera acquiring the initial chip image should be sufficient to clearly distinguish the smallest particles of the frosted texture, to avoid failing to capture texture details, which would result in the absence of corresponding mid-frequency components in the spectrogram, meaning the frosted defects would be undetectable. Simultaneously, the camera's viewing distance should be controlled to prevent the background from occupying too large a proportion in the initial chip image, while the area occupied by the frosted defects is too small, leading to a minimal contribution of the defect to the mid-frequency energy, and the subsequent energy ratio being dominated by the background energy. This embodiment of the invention does not impose limitations on this aspect.

[0030] The main characteristic of frosted defects is abnormal micro-roughness of the semiconductor chip surface. Normal chip surfaces exhibit roughness at specific frequencies, while the amplitude of the mid-frequency components increases in frosted defect regions. Under the same illumination, the average gray value may increase due to enhanced scattering or decrease due to absorption, and the local variance of gray value differs from that of normal regions. This defect lacks a clear edge contour; instead of a gradual transition from the normal region, the distribution of high-frequency textures and the pattern of gray-level variation on the surface are altered.

[0031] Step 102: Obtain the spectrum diagram based on the initial chip image.

[0032] In the initial chip image, frosted defects exhibit slight grayscale variations. However, the grayscale distribution is simultaneously affected by both illumination gradients and the actual texture. Furthermore, camera sensor noise and photon noise in the spatial domain also manifest as random grayscale fluctuations, closely resembling the visual effect of frosted textures. This leads to missed or false detections of frosted defects during defect detection. Spectral analysis, on the other hand, can distinguish mid-frequency frosted defects from high-frequency noise and low-frequency illumination gradients, making it more suitable for detecting frosted defects than simply analyzing grayscale distribution.

[0033] In this embodiment of the invention, the spectrum diagram can be used to represent image frequency and energy distribution. The method of obtaining the spectrum diagram based on the initial chip image may include: I. Wavelet Transform Specifically, it can be either continuous wavelet transform (CWT) or discrete wavelet transform (DWT). For example, by analyzing the initial chip image using two-dimensional DWT, sub-bands of low-frequency approximation, horizontal detail, vertical detail, and diagonal detail can be obtained. Each sub-band is a function of spatial coordinates, reflecting the energy distribution at different scales and directions.

[0034] II. Hilbert-Huang Transform (HHT) First, the signal is decomposed into several intrinsic mode functions (IMFs) using Empirical Mode Decomposition (EMD). Then, a Hilbert transform is performed on each IMF. In the initial chip image, this is represented by scanning the image with a moving window, calculating the local energy spectrum within each window. On the time-frequency plane, each time point corresponds to an instantaneous frequency. The energies of all IMFs are superimposed according to time and frequency to obtain the Hilbert spectrum. The energy intensity in this spectrum is represented by instantaneous amplitude / square of amplitude, and can be represented by color in the graph, with a very fine energy distribution.

[0035] III. Fourier Transform Preferably, a two-dimensional discrete Fourier transform (DFT) can be used. The radial direction (distance of a pixel from the center of the spectrogram) represents the frequency magnitude. The closer to the center, the lower the frequency, corresponding to slowly changing backgrounds and large areas in the image; the closer to the center, the higher the frequency, corresponding to details, textures, edges, noise, etc. in the image. The horizontal axis of the spectrogram is the horizontal frequency, and the vertical axis is the vertical frequency. The grayscale / brightness on the spectrogram is used to represent the energy intensity of the corresponding frequency component; bright white indicates strong frequency energy, and dark black indicates weak frequency energy.

[0036] Furthermore, the spectrum can also be obtained using the Fast Fourier Transform (FFT), which can output fine frequency resolution from a global perspective, and the energy integral is direct and has a clear physical meaning. Alternatively, a spectrum that can be indirectly obtained using Discrete Cosine Transform (DCT), Gabor Transform, etc., can be used to represent frequency and energy distribution; however, this embodiment of the invention does not limit the specific methods used.

[0037] Step 103: Divide the spectrum into grids to obtain multiple grid cells; select candidate grid cells; select the first grid cell.

[0038] In this embodiment of the invention, the shape and size of the grid cells are not limited. Preferably, a frequency resolution-based partitioning method can be used, with the basic unit corresponding to the frequency resolution as the size of the grid cell. The larger the initial chip image size, the higher the frequency resolution, and thus the richer the distinguishable frequency details, resulting in sparser grid cells. Larger grid cells can also be used. When a grid cell contains multiple spectral pixels, a superpixel method can be used to condense the image information. This is not limited in this embodiment of the invention.

[0039] Since frosted defects typically appear as patches of light that continuously decay outwards from the center in a spectrum, these patches are diffuse, roughly circular, bright areas. These patches darken continuously outwards from the center without a clear edge. Therefore, in this embodiment of the invention, the candidate mesh unit is a mesh unit whose energy is greater than a preset first threshold corresponding to an adjacent mesh unit. The adjacent mesh unit can be a mesh unit adjacent to the candidate mesh unit. This adjacency can be represented by edge contact, or by defining a nearest neighbor radius or number. If the mesh is not divided according to the label, or uses discrete points, the adjacent mesh units can be determined based on the radius and the number of nearest neighbors. This embodiment of the invention does not impose such limitations.

[0040] In this embodiment of the invention, the purpose of setting a first threshold to restrict adjacent grid cells is to restrict the edges of the area composed of all candidate grid cells, i.e., the light spot that may correspond to the frosted defect, based on the diffuse nature of the frosted defect in the spectrum. Since this judgment condition is set on adjacent grid cells, the outermost layer of the grid in this area will have an additional layer of candidate grid cells with actual energy less than the preset first threshold, so as to avoid the edge restriction causing more energy loss in the case of a large area of ​​light spot.

[0041] In this embodiment of the invention, the first grid cell can be a candidate grid cell whose frequency is greater than a preset second threshold and less than a preset third threshold. The significance of setting the second threshold is to delineate the boundary between the first and second frequency bands in the mid-to-low frequency range. Firstly, due to the characteristics of the frosted defect in the spectrum, its frequency range is 0-X. Therefore, if a hard boundary is directly delineated based on an artificially set threshold, the resulting two frequency bands may be inaccurate, thus affecting the accuracy of the energy ratio. Therefore, the two frequency bands are only roughly delineated at the second threshold, which can be used to constrain the boundary of the two frequency bands, with the aim of making their frequency range near the second threshold. After the boundary is determined by the second threshold, the distinguishing element of the frosted defect can be separated from the background frequency, thereby making the ratio relationship between the two frequency bands clearer.

[0042] The third threshold is set to distinguish the core area, defects with obvious edges, and other high-frequency noise in the image. The first grid cell is then determined based on the range defined by the second and third thresholds. The third threshold is higher than the second threshold.

[0043] Step 104: Obtain the first frequency band based on the first grid cell; obtain the second frequency band based on the spectrum and the second threshold.

[0044] In this embodiment of the invention, a frequency band can be a frequency range, a selection box corresponding to that frequency range, or a region on a spectrum graph. For example, in a two-dimensional discrete Fourier transform, a frequency band can be defined as a frequency range. In this case, all the energy corresponding to the frequency band mentioned in subsequent steps is the total energy of each frequency branch within the frequency range. A frequency band can also be defined as a selection box corresponding to a frequency range. For example, in this embodiment of the invention, for a medium frequency such as a frosted defect (not low-frequency background, not high-frequency noise), the frequency band is represented as a circular selection box in the spectrum graph obtained by the Fourier transform, and the energy corresponding to the frequency band is the sum of the energy corresponding to all frequency branches within the selection box. When the spectrum is defined as a region, it can be understood as the selected area on the spectrum graph corresponding to that selection box, which will not be elaborated further in this embodiment of the invention.

[0045] The first frequency band can be used to represent the predicted frequency range of frosted defects in the initial chip image. Specifically, since the frequency range of frosted defects is 0-X, while what we actually want is the mid-frequency range of YX (Y>0), which corresponds to the distinguishing features between frosted defects and the background, the first frequency band can be used to represent the predicted frequency range of distinguishing features in the initial chip image. These distinguishing features can be obtained based on the predicted frequency range of the frosted defects and the corresponding predicted frequency range of the background in the initial chip image.

[0046] In this embodiment of the invention, the process of obtaining the first frequency band based on the first grid cell can specifically involve obtaining the predicted frequency range based on the minimum and maximum frequencies corresponding to multiple contiguous regions formed by all the first grid cells, or determining the approximate boundary based on the discreteness / distribution of the first grid cells. While the distribution can effectively eliminate some interference, it also narrows the frequency band range. This embodiment of the invention does not limit this approach. Since the predicted frequency range is defined based on the frequency range of the distinguishing characteristics of the frosted defect, the predicted frequency range can be an adaptive fine-tuning based on the first grid cells. This embodiment of the invention does not limit this approach either.

[0047] Furthermore, the process of obtaining the frequency band based on the grid cells can also involve extracting the frequency band separately from a large area with independent extreme frequency ranges. Therefore, the first frequency band can also include multiple discrete frequency bands, and this embodiment of the invention does not limit this. The purpose of setting a first frequency band instead of directly using the first grid cells is that the first grid cells can reflect the real continuous spectral points in the spectrum, while the first frequency band is a predicted range obtained based on these spectral points. This prediction process can be based on machine learning methods, which train a model to identify the spectral characteristics of the frosted defects and directly give the corresponding first frequency band based on the first grid cells. Alternatively, a Bayesian update method can be used, which also requires data as prior or likelihood estimation samples, and then predicts the range corresponding to the frequency band based on historical data or sample data. This embodiment of the invention does not limit this.

[0048] The second frequency band can be used to represent the predicted frequency range of the overall brightness information of the initial chip image. The setting method of the second frequency band is the same as that of the first frequency band, and will not be repeated here. The second frequency band corresponds to the energy in the low-frequency or extremely low-frequency portion, which corresponds to the background and smooth interface in the initial chip image.

[0049] Step 105: Calculate the ratio of all energies corresponding to the first frequency band to all energies corresponding to the second frequency band to obtain the energy ratio.

[0050] In this embodiment of the invention, the energy ratio is used to represent the degree of frosting defects in the initial chip image. The energy in the first frequency band primarily reflects the predicted energy of the frosting defects, which increases significantly with increasing chip surface roughness. The energy in the second frequency band primarily reflects the overall brightness of the image and the reflection from the smooth surface of the chip. When differences arise in different images due to uneven illumination, variations in light intensity, camera exposure time, or local shadows, the energies of the two frequency bands are scaled by almost the same factor because both originate from a linear transformation of the same initial chip image. Therefore, comparing these two energies can offset the aforementioned scaling effects or differences, making the energy ratio a wear intensity indicator independent of lighting conditions.

[0051] As can be seen, in this embodiment of the invention, firstly, the initial chip image is converted into a spectrum diagram corresponding to the chip through Fourier transform, fully considering the characteristics of frosted defects being non-high-frequency and edgeless. Furthermore, the first grid cell determined through two judgment conditions can be used to represent the coverage area of ​​the frosted defect in the spectrum diagram. Then, based on all the first grid cells, a first frequency band is obtained that can represent the predicted frequency range of the frosted defect in the initial image. Since the second frequency band corresponds to the low-frequency range and is used to represent the overall brightness, and also includes low-frequency information corresponding to smooth interfaces, finally, by comparing all the energies within the two frequency bands, the energy ratio can accurately represent the intensity of the frosted defect, independent of illumination conditions, based on fully considering the characteristics of the frosted defect, thereby improving the accuracy of semiconductor chip appearance defect detection.

[0052] In an optional embodiment, to specify the method of obtaining the first frequency band, the above-described method of obtaining the first frequency band based on the first grid cell may further include: The initial frequency band is obtained based on the first grid cell; Perform an adjacency update operation on the initial frequency band: construct the adjacent frequency band; update the initial frequency band based on the adjacent frequency band to obtain the updated initial frequency band and the first adjustment parameter; When it is determined that the first adjustment parameter is less than the preset fourth threshold, the updated initial frequency band is determined to be the first frequency band.

[0053] In this optional embodiment, the initial frequency band can be used to represent the frequency range corresponding to the first grid cell. The adjacent frequency band is adjacent to the initial frequency band. The adjacent frequency band can be the frequency band form described in the above embodiment, or it can be a single frequency value or a region corresponding to that frequency. For example, in the spectrum obtained by Fourier transform, when the adjacent frequency band corresponds to the region represented by a single frequency value, it can present a circle. In the subsequent process of updating the initial frequency band, if the adjacent frequency band is a frequency value and the initial frequency band is a frequency range or a selection box, it is reflected as the expansion of the frequency range or the expansion of the selection box boundary. If both the adjacent frequency band and the frequency range are regions, it is the splicing of regions or the splicing of the content in the corresponding spectrum.

[0054] In this optional embodiment, the first adjustment parameter is used to measure the change in dispersion before and after the initial frequency band update. Dispersion is used to evaluate the degree of dispersion of the energy distribution corresponding to the frequency band. This index can be variance, Gini coefficient, JS divergence, etc. It is mainly used to reflect whether the energy distribution is discrete, which may be reflected in the spectrum as the degree of dispersion of gray-level distribution. In the spectrum obtained by Fourier transform, the degree of dispersion / concentration of the light spot is essentially the degree of energy diffusion in the radial direction. Rough texture will cause energy to diffuse from the center to the periphery, and the high-frequency part will show a halo-like or radial tail. Therefore, the core part of the frosted defect is represented by a disk that is brightest in the center and gradually darkens to the periphery after the two-dimensional discrete Fourier transform. Moreover, the boundary cannot be determined in the spectrum. Therefore, after defining the frequency corresponding to the frosted defect as the intermediate frequency, the boundary between the intermediate frequency and the high frequency can be regarded as the process of finding adjacent frequency bands in this optional embodiment as a dynamic process of continuously extending outwards based on the initial frequency band. This process is highly consistent with the diffusion characteristics of the frosted defect in the spectrum. And this process of extending outwards is to find a suitable dividing boundary.

[0055] Due to the gradual fading of the frosted defects, the corresponding spectral point energy in the spectrum becomes increasingly lower and sparser. Therefore, this increasingly sparse appearance can be measured by the degree of dispersion, resulting in an increasingly discrete appearance. This optional embodiment does not limit this.

[0056] It should be noted that the fourth threshold can be zero, meaning it is the boundary point between the concentrated and discrete states, which perfectly matches the variation characteristics of the frosted defect in the spectrum. Since the frequency range covered by the frosted defect is an interval starting from 0, the second frequency band corresponding to the low frequency overlaps with the first frequency band. When distinguishing the boundary of the mid-low frequency band, a lower fourth threshold than that for the mid-high frequency band can be used to limit the rapid convergence of the first adjustment parameter.

[0057] As can be seen, this optional embodiment can further refine the confirmation process of the first frequency band through the adjacent update operation, so that the initial frequency band is updated to the point where the energy concentration changes, which is consistent with the diffusion and gradual change characteristics of the frosted defect in the spectrum, thereby improving the acquisition accuracy of the frequency band corresponding to the frosted defect, and thus improving the accuracy of the frosted defect detection process.

[0058] In another alternative embodiment, the method may further include: Based on the updated initial frequency band and adjacent frequency bands, the adjacency coordination parameters are obtained. The adjacency coordination parameters are used to measure the change in the dispersion between the adjacent frequency bands and the updated initial frequency bands. Based on the adjacency coordination parameters and the first adjustment parameters, the second adjustment parameters are obtained. The second adjustment parameters are used to measure the synchronization between the adjacency coordination parameters and the first adjustment parameters. When it is determined that the second adjustment parameter does not meet the preset frequency band setting conditions, multiple rounds of adjacency update operations are performed on the updated initial frequency band to determine the first frequency band based on the update results.

[0059] In this optional embodiment, the purpose of setting the adjacency coordination parameter is to provide a reverse buffer reference for the initial frequency band update process: the adjacency coordination parameter measures the change in the concentration of adjacent frequency bands in the opposite direction of the initial frequency band update.

[0060] It should be noted that the introduction of adjacency coordination parameters is mainly aimed at the first frequency band. When the boundary between the mid- and low-frequency bands is difficult to determine, a reverse asynchronous confirmation method is adopted. Taking the mid- and high-frequency boundary of the first frequency band as an example, during the continuous updating of the first frequency band with adjacent frequency bands, the first adjustment parameter reflects the increasingly dispersed nature of the first frequency band. However, due to the overall dispersion, the adjacency coordination parameters show a concentrated trend during the update process from adjacent frequency bands to the updated initial frequency band. This asynchronous situation in the reverse process is reflected in the second adjustment parameter as asynchronous (low synchronization), thus supporting further updates of the initial frequency band.

[0061] At the boundary between the low and mid-frequency bands, updates are usually performed on the original frequency band. Introducing new adjacent frequency bands may cause the overall process to become increasingly discrete. This makes it difficult for the initial frequency band update boundary process to converge when relying solely on the first adjustment parameter. The process from the low frequency band to the mid frequency band, i.e., from the adjacent frequency band to the updated initial frequency band, is also discrete. At this point, by using the adjacent coordination parameter, the second adjustment parameter exhibits high synchronization, thus determining that the preset frequency band establishment conditions are not met and stopping the update. This effectively prevents the first frequency band from over-merging during the update process.

[0062] The synchronization status represented by the second adjustment parameter is reflected in the above process. Specifically, synchronization can be a reflection of consistency or similarity. Therefore, the second adjustment parameter can be the Pearson correlation coefficient, JS divergence, KL divergence, or consistency correlation coefficient (CCC). This optional embodiment does not limit this.

[0063] In this optional embodiment, when the frequency band setting conditions are met, it is determined that the frequency band can be used for subsequent ratio calculations. The purpose of using the frequency band setting conditions is to constrain the second adjustment parameter, prevent overfitting, and make adaptive adjustments for different images based on image size, resolution, etc.

[0064] As can be seen, in this optional embodiment, the adjacency coordination parameter provides reverse confirmation, thereby enabling the acquisition accuracy of the first frequency band to be gradually improved in areas where it is difficult to define boundaries in the mid-to-low frequency range. At the same time, the preset frequency band setting conditions prevent underfitting and overfitting in the update process, thereby improving the robustness of the update process and thus improving the accuracy of the abrasive defect detection.

[0065] In yet another optional embodiment, the method may further include: Construct secondary adjacent frequency bands, which are frequency bands that are adjacent to adjacent frequency bands in directions other than the initial frequency band; Based on the secondary adjacent frequency band, the secondary coordination parameters are obtained. The secondary coordination parameters are used to measure the change in the dispersion between the adjacent frequency band and the secondary adjacent frequency band. Furthermore, the secondary coordination parameters are also used to: when it is determined that the second adjustment parameter does not meet the preset frequency band setting conditions, control the upper limit of the adjacent update rounds to be lower than the preset fifth threshold.

[0066] In this optional embodiment, the secondary adjacent frequency band can be acquired using the same method as the adjacent frequency band, and this optional embodiment will not be elaborated further. The purpose of designing the secondary adjacent frequency band is to reflect the energy distribution change trend and direction between adjacent frequency bands through secondary coordination parameters. The secondary adjacent frequency band can also be presented as a stepped change in the degree of dispersion using multiple parameters, thereby providing a reference for whether the first frequency band needs to be updated / expanded. The first adjustment parameter is set to reflect the energy distribution change trend and direction between the first frequency band and the adjacent frequency band. Therefore, the process of controlling the update cycle can also be achieved by comparing the secondary coordination parameters with the first adjustment parameters. By studying the update amplitude of the first frequency band and the update amplitude of the adjacent frequency band, the convergence prediction of this dispersion case during the update process can be determined, thereby further improving the accuracy of the first frequency band.

[0067] In this optional embodiment, the purpose of using the frequency band setting condition is to constrain the second adjustment parameter to prevent overfitting, and to make adaptive adjustments for different images based on image size, resolution, etc. The secondary coordination parameter is used to limit overfitting at the round level.

[0068] As can be seen, this optional embodiment can determine the convergence of the initial frequency band update process through the secondary coordination parameters and the first coordination parameters, thereby controlling the number of adjacent update rounds in advance to prevent overfitting and thus obtaining a more accurate first frequency band, improving the accuracy of semiconductor chip appearance defect detection.

[0069] In yet another optional embodiment, the method may further include: The spectrogram is inversely transformed to obtain the inverse graph, which is used to represent the corresponding positional relationship of image frequencies on the initial chip image. The inverter diagram is divided into grids to select the second grid cells, where the second grid cells are the grid cells corresponding to the first frequency band in the inverter diagram; Based on the second grid cell, a confidence parameter is obtained, which is used to measure the confidence that the second grid cell contains a frosted defect. Based on the confidence parameter, all energies corresponding to the first frequency band are updated to obtain the updated energy ratio, which is used to represent the degree of frosted defect in the initial chip image.

[0070] In this optional embodiment, the spectrogram is derived from the initial chip image in the spatial domain. However, relying solely on this image for detection and separation results in the loss of all spatial location information, making it impossible to locate defects. Furthermore, it cannot distinguish between signals with similar frequencies but different origins. The inverse transformation, converting the spectrogram back from the frequency domain to the spatial domain, allows global frequency information to be reassigned to the spatial domain, thus revealing the correspondence between frequency domain energy, frequency, and the location of the original image.

[0071] The reason for setting the confidence level parameter is that the size of the second grid cell may contain multiple pixels, and different pixels may represent different image information. Preferably, the size of the second grid cell can be adaptively adjusted according to the resolution, so that each grid covers a single pixel. Another reason for setting the confidence level is that the same pixel may also have occlusion or overlap in real-world scenarios, such as intersecting with functional areas or overlapping with other types of defects. Therefore, it is difficult to measure whether the frequency corresponding to the frosted defect at that point provides frequency domain energy.

[0072] The confidence parameter can be a probability value based on the model's output. This model can be a deep semantic segmentation network model such as DeepLab, U-Net, or SegFormer, where the Softmax function is used in the last layer of the network to output the probability distribution of all categories. The confidence parameter can also be an uncertainty measure such as prediction entropy or marginal confidence; this optional embodiment is not limited in this regard.

[0073] As can be seen, this optional embodiment can further improve the accuracy of abrasive defect detection by performing confidence analysis from the perspective of the spatial domain and based on the data obtained from the inverse transformation process through the inverse transformation operation.

[0074] In yet another optional embodiment, the method may further include: Obtain a sample spectrum atlas, which includes multiple sample spectrum images. The sample spectrum images are the spectrum images corresponding to semiconductor chip images with frosted defects. Based on the trained detection model, the spectrogram is analyzed to obtain the third frequency band; the trained detection model is obtained by training on a sample spectrogram set. Based on the first frequency band and the third frequency band, a consistent coordination parameter is obtained. The consistent coordination parameter is used to measure the consistency between the first frequency band and the third frequency band. The consistent coordination parameter is used to control the upper limit of the adjacent update round to be lower than the fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band setting conditions.

[0075] In this optional embodiment, the sample frequency domain atlas can be a collection of sample images used for training the detection model. Therefore, the corresponding sample images preferably have frosted defects so that the model can assign a suitable frequency band in the mid-frequency band to the frosted region based on the spectrogram after training.

[0076] In this optional embodiment, the detection model can be a SimpleNet model, or a powerful teacher network like SuperSimpleNet that is pre-trained with samples, followed by training a lightweight student network to mimic the teacher's output during the detection process. Alternatively, it can be a CNN-based method, such as PatchCore, which generates a high-quality image feature memory on the spectrogram and determines the frequency band of the frosted defect by calculating the nearest neighbor distance between the spectrogram features and the memory. Another option is a NexViTAD-based method, which can solve the domain transfer problem in real-world scenarios through multi-task learning and shared subspace projection.

[0077] It should be noted that neither the frequency band obtained by setting a hard threshold manually nor the frequency band used in the calculation process based on the model output can guarantee accuracy. This is because the frosted defect is represented in the spectrum as a frequency band starting from 0. Therefore, it is destined to be mixed with low-frequency information, making it difficult to determine a clear boundary with practical significance. Only the result of further refining the frequency band energy ratio can be obtained.

[0078] As can be seen, in order to improve the accuracy and universality of the energy ratio calculation process, this optional embodiment adopts a soft-binding method between the third frequency band obtained by the model and the first frequency band to judge their consistency / update completion, which is used to further constrain the upper limit of the adjacent update rounds and prevent underfitting or overfitting.

[0079] In yet another optional embodiment, the method may further include: Based on the current adjacency update round, the fourth threshold is updated to narrow the judgment range of the first adjustment parameter; Based on the current adjacency update round, the adjacent frequency bands are updated to narrow down the frequency range corresponding to the adjacent frequency bands.

[0080] In this optional embodiment, the fourth threshold is a threshold that restricts the judgment condition of the first adjustment parameter. The update operation of setting the fourth threshold is essentially to constrain the update rounds. At the same time, it is also to quickly establish the mid-to-low frequency band boundary in the spectrum. During the process of updating from mid-frequency to low-frequency adjacency, the first adjustment parameter may always remain in an increasingly discrete state. At this time, adding this condition can effectively prevent it from merging. In addition, the means of updating the fourth threshold can also serve as a simplification / replacement operation for judging the second adjustment parameter, which facilitates the faster completion of the initial frequency band update in large-scale data processing, thereby quickly obtaining the range of the first frequency band.

[0081] In this optional embodiment, updating the adjacent frequency band is to improve the accuracy of the first frequency band in determining its boundaries. This is because, initially, the width of the adjacent frequency band might be set too wide. However, as the adjacent frequency band update operation progresses, an excessively wide adjacent frequency band can cause the update process to lose a large amount of discrete information or obscure the changes in the degree of dispersion in the parameters. Consequently, the boundary of the obtained first frequency band becomes inaccurate, and the frequency domain energy ratio cannot effectively reflect the degree and state of the frosting defect. Therefore, it is necessary to gradually narrow the boundary according to the update rounds to achieve refined operation.

[0082] As can be seen, this optional embodiment, by updating the fourth threshold and adjacent frequency bands through the update rounds, can further restrict the update judgment conditions and achieve a reasonable allocation of refined operation resources, thereby improving the accuracy of the first frequency band and thus improving the accuracy of grinding defect detection.

[0083] It should be noted that, in the above embodiments, the spectrum references not illustrated can all be obtained by using the spectrum obtained by the two-dimensional discrete Fourier transform process to implement the technical means in the above embodiments. In this regard, the embodiments of the present invention do not limit the scope of the references.

[0084] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a semiconductor chip appearance defect detection device disclosed in an embodiment of the present invention, as shown below. Figure 2 As shown, the semiconductor chip appearance defect detection device may include: The image acquisition module 201 is used to acquire an initial chip image; based on the initial chip image, a spectrum is obtained, which is used to represent the image frequency and energy distribution; The grid filtering module 202 is used to divide the spectrum into grids to obtain multiple grid cells; filter to obtain candidate grid cells, wherein the candidate grid cells are grid cells whose energy corresponding to adjacent grid cells is greater than a preset first threshold; filter to obtain a first grid cell, wherein the first grid cell is a candidate grid cell whose corresponding frequency is greater than a preset second threshold and less than a preset third threshold; The frequency band confirmation module 203 is used to obtain a first frequency band based on the first grid cell; the first frequency band is used to represent the predicted frequency range of the frosted defects in the initial chip image; and to obtain a second frequency band based on the spectrum and the second threshold, the second frequency band being used to represent the predicted frequency range of the overall brightness information of the initial chip image. The energy ratio module 204 is used to calculate the ratio of all energies corresponding to the first frequency band to all energies corresponding to the second frequency band to obtain the energy ratio. The energy ratio is used to represent the degree of frosting defects in the initial chip image.

[0085] In an optional embodiment, the frequency band confirmation module 203 obtains the specific operation mode of the first frequency band based on the first grid cell, which may include: Based on the first grid cell, an initial frequency band is obtained, which is used to represent the frequency range corresponding to the first grid cell; Perform an adjacency update operation on the initial frequency band: construct the adjacency frequency band; based on the adjacency frequency band, update the initial frequency band to obtain the updated initial frequency band and the first adjustment parameter, wherein the first adjustment parameter is used to measure the change in the dispersion of the initial frequency band before and after the update, and the dispersion is used to evaluate the degree of dispersion of the energy distribution corresponding to the frequency band; When it is determined that the first adjustment parameter is less than the preset fourth threshold, the updated initial frequency band is determined to be the first frequency band.

[0086] In another alternative embodiment, the device may further include: The synchronization adjustment module is used to obtain adjacency coordination parameters based on the updated initial frequency band and adjacent frequency bands. The adjacency coordination parameters are used to measure the change in the dispersion between the adjacent frequency band and the updated initial frequency band. Based on the adjacency coordination parameters and the first adjustment parameters, a second adjustment parameter is obtained. The second adjustment parameter is used to measure the synchronization between the adjacency coordination parameters and the first adjustment parameter. When it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, multiple rounds of adjacency update operations are performed on the updated initial frequency band to determine the first frequency band based on the update results.

[0087] In yet another alternative embodiment, the device may further include: The secondary coordination module is used to construct secondary adjacent frequency bands, which are frequency bands that are adjacent to each other in directions other than the initial frequency band. Based on the secondary adjacent frequency bands, secondary coordination parameters are obtained. The secondary coordination parameters are used to measure the change in the dispersion between adjacent frequency bands and secondary adjacent frequency bands. In addition, the secondary coordination parameters are also used to control the upper limit of the number of adjacent update rounds to be lower than the preset fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions.

[0088] In yet another alternative embodiment, the device may further include: The inverter detection module performs an inverse transformation on the spectrogram to obtain an inverter map, which represents the corresponding positional relationship of image frequencies on the initial chip image. The inverter map is then divided into grids to filter out second grid cells, where the second grid cell is the grid cell corresponding to the first frequency band in the inverter map. Based on the second grid cell, a confidence parameter is obtained, which measures the confidence level that the second grid cell contains a frosted defect. Based on the confidence parameter, all energies corresponding to the first frequency band are updated to obtain an updated energy ratio, which represents the degree of frosted defect in the initial chip image.

[0089] In yet another alternative embodiment, the device may further include: The sample detection module acquires a sample spectrum set, which includes multiple sample spectrum images corresponding to semiconductor chip images with frosted defects. Based on a trained detection model, the spectrum images are analyzed to obtain a third frequency band. The trained detection model is obtained by training based on the sample spectrum set. Based on the first and third frequency bands, a consistency coordination parameter is obtained. The consistency coordination parameter is used to measure the consistency between the first and third frequency bands. The consistency coordination parameter is used to control the upper limit of the adjacent update rounds to be lower than the fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band setting conditions.

[0090] In yet another alternative embodiment, the device may further include: The range reduction module is used to update the fourth threshold based on the current adjacency update round to narrow the judgment range of the first adjustment parameter; and to update the adjacent frequency band based on the current adjacency update round to narrow the frequency range corresponding to the adjacent frequency band.

[0091] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a semiconductor chip appearance defect detection system disclosed in an embodiment of the present invention. Figure 3 As shown, the semiconductor chip appearance defect detection system may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in any of the semiconductor chip appearance defect detection methods in Embodiment 1 of the present invention.

[0092] Example 4 This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute some or all of the steps in any of the semiconductor chip appearance defect detection methods disclosed in Embodiment 1 of this invention.

[0093] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0094] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0095] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting surface defects in semiconductor chips, characterized in that, The method includes: Obtain an initial chip image; based on the initial chip image, obtain a spectrum diagram, which is used to represent the image frequency and energy distribution; The spectrum is divided into grids to obtain multiple grid cells; Candidate grid cells are obtained through screening, wherein the candidate grid cells are grid cells whose energy is greater than a preset first threshold corresponding to adjacent grid cells; The first grid cell is obtained by filtering, wherein the first grid cell is a candidate grid cell whose frequency is greater than a preset second threshold and less than a preset third threshold; A first frequency band is obtained based on the first grid cell; the first frequency band is used to represent the predicted frequency range of frosted defects in the initial chip image; a second frequency band is obtained based on the spectrum and the second threshold; the second frequency band is used to represent the predicted frequency range of the overall brightness information of the initial chip image. The energy ratio is obtained by comparing all the energies corresponding to the first frequency band with all the energies corresponding to the second frequency band. The energy ratio is used to represent the degree of frosting defects in the initial chip image.

2. The semiconductor chip appearance defect detection method according to claim 1, characterized in that, The step of obtaining the first frequency band based on the first grid cell further includes: Based on the first grid cell, an initial frequency band is obtained, which is used to represent the frequency range corresponding to the first grid cell; The initial frequency band is subjected to an adjacency update operation: an adjacency frequency band is constructed; based on the adjacency frequency band, the initial frequency band is updated to obtain the updated initial frequency band and a first adjustment parameter, wherein the first adjustment parameter is used to measure the change in the dispersion of the initial frequency band before and after the update, and the dispersion is used to evaluate the degree of dispersion of the energy distribution corresponding to the frequency band; When it is determined that the first adjustment parameter is less than the preset fourth threshold, the updated initial frequency band is determined to be the first frequency band.

3. The semiconductor chip appearance defect detection method according to claim 2, characterized in that, The method further includes: Based on the updated initial frequency band and the adjacent frequency band, an adjacency coordination parameter is obtained. The adjacency coordination parameter is used to measure the change in the dispersion between the adjacent frequency band and the updated initial frequency band. A second adjustment parameter is obtained based on the adjacency coordination parameter and the first adjustment parameter. The second adjustment parameter is used to measure the synchronization between the adjacency coordination parameter and the first adjustment parameter. When it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, multiple rounds of the adjacency update operation are performed on the updated initial frequency band to determine the first frequency band based on the update results.

4. The semiconductor chip appearance defect detection method according to claim 3, characterized in that, The method further includes: Construct a secondary adjacent frequency band, wherein the secondary adjacent frequency band is a frequency band that is adjacent to the adjacent frequency band in a direction other than the initial frequency band; Based on the secondary adjacent frequency band, a secondary coordination parameter is obtained. The secondary coordination parameter is used to measure the change in the dispersion between the adjacent frequency band and the secondary adjacent frequency band. Furthermore, the secondary coordination parameter is also used to: when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions, control the upper limit of the adjacency update round to be lower than the preset fifth threshold.

5. The semiconductor chip appearance defect detection method according to claim 4, characterized in that, The method further includes: The spectrum is inversely transformed to obtain an inverse graph, which is used to represent the corresponding positional relationship of image frequencies on the initial chip image; The inverter diagram is divided into grids to select a second grid cell, wherein the second grid cell is the grid cell corresponding to the first frequency band in the inverter diagram; Based on the second grid cell, a confidence parameter is obtained, which is used to measure the confidence that the second grid cell contains a frosted defect; based on the confidence parameter, all energies corresponding to the first frequency band are updated to obtain an updated energy ratio, which is used to represent the degree of frosted defect in the initial chip image.

6. The semiconductor chip appearance defect detection method according to claim 5, characterized in that, The method further includes: Obtain a sample spectrum set, which includes multiple sample spectrum images, wherein the sample spectrum images are spectrum images corresponding to semiconductor chip images with frosted defects; Based on the trained detection model, the spectrogram is analyzed to obtain the third frequency band; wherein, the trained detection model is trained based on the sample spectrogram set; Based on the first frequency band and the third frequency band, a consistent coordination parameter is obtained. The consistent coordination parameter is used to measure the consistency between the first frequency band and the third frequency band. The consistent coordination parameter is used to control the upper limit of the adjacent update round to be lower than the fifth threshold when it is determined that the second adjustment parameter does not meet the preset frequency band establishment conditions.

7. The semiconductor chip appearance defect detection method according to claim 6, characterized in that, The method further includes: Based on the current adjacency update round, the fourth threshold is updated to narrow the judgment range of the first adjustment parameter; Based on the current adjacency update round, the adjacent frequency band is updated to narrow the frequency range corresponding to the adjacent frequency band.

8. A semiconductor chip appearance defect detection device, characterized in that, The device includes: An image acquisition module is used to acquire an initial chip image; based on the initial chip image, a spectrum diagram is obtained, which is used to represent the image frequency and energy distribution; A grid filtering module is used to divide the spectrum into grids to obtain multiple grid cells; filter to obtain candidate grid cells, wherein the candidate grid cells are grid cells whose energy corresponding to adjacent grid cells is greater than a preset first threshold; filter to obtain a first grid cell, wherein the first grid cell is a candidate grid cell whose corresponding frequency is greater than a preset second threshold and less than a preset third threshold. A frequency band confirmation module is used to obtain a first frequency band based on the first grid cell; the first frequency band is used to represent the predicted frequency range of frosted defects in the initial chip image; and to obtain a second frequency band based on the spectrum and the second threshold, the second frequency band being used to represent the predicted frequency range of the overall brightness information of the initial chip image. An energy ratio module is used to calculate the ratio of all energies corresponding to the first frequency band to all energies corresponding to the second frequency band, thereby obtaining an energy ratio. The energy ratio is used to represent the degree of frosting defects in the initial chip image.

9. A semiconductor chip appearance defect detection system, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the semiconductor chip appearance defect detection method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked by a processor, execute the semiconductor chip appearance defect detection method as described in any one of claims 1-7.

Images