Detecting defects in semiconductor samples using weak labelling

Abstract

To provide a system for classifying an interest pattern (POI) on a semiconductor specimen.SOLUTION: A system comprises: a processor; and a memory circuit. The processor and the memory circuit, acquires a high resolution image of POI, and is structured so as to generate data usable for classifying the POI in accordance with a failure related classification. In the generation, a machine learning model trained in accordance with a training sample is used. The training sample is a high resolution training image captured by scanning each training pattern on a sample, and each training patter includes the high resolution training image which is similar to the POI, and a labeling related to an image. The labeling is derived from a low resolution inspection of each training pattern.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The subject matter disclosed herein relates generally to the field of testing of wafer specimens, and more particularly to detecting defects within the specimens. [Background technology]

[0002] Presently, the increased density and performance associated with ultra-large scale integration of manufactured devices requires sub-micron features, increased transistor and circuit speeds, and improved reliability. Such demands necessitate the formation of device features with high precision and uniformity, which in turn necessitates careful oversight of the manufacturing process, including automatic testing of the devices while still in semiconductor wafer form.

[0003] As a non-limiting example, runtime testing may use a two-stage procedure, for example, inspecting a sample followed by examining potential defect locations. During the first stage, the surface of the sample is inspected at a high speed and at a relatively low resolution. In the first stage, a defect map is created to indicate suspect locations on the sample that have a high probability of defects. During the second stage, at least some of the suspect locations are more thoroughly analyzed at a relatively high resolution. In some cases, both stages may be performed by the same inspection tool, and in some other cases, the two stages are performed by different inspection tools.

[0004] At various steps during semiconductor manufacturing, testing processes are used to detect and classify defects on the specimens. The effectiveness of the testing can be increased by automating the process, e.g., automatic defect classification (ADC), automatic defect review (ADR), etc. Summary of the Invention

[0005] According to one aspect of the subject matter disclosed herein, there is provided a system for classifying a pattern of interest (POI) on a semiconductor specimen, the system comprising a processor and memory circuit (PMC), the PMC comprising: acquiring data indicative of a high-resolution image of the POI on the sample; configured to generate data usable to classify the POI according to a defect-related classification; The generation utilizes a machine learning model trained according to at least a plurality of training samples, each of which comprises: high-resolution training images captured by scanning a respective training pattern on the sample, each training pattern being similar to a POI; Each training pattern is associated with a label that is derived from a low-resolution inspection of the respective training pattern.

[0006] In addition to the features described above, systems according to this aspect of the presently disclosed subject matter may include one or more of the following features (i)-(xiii) in any desired combination or permutation technically possible: (i) the high-resolution training image of each training sample of the plurality of training samples is a scanning electron microscope (SEM) image; (ii) a label for each training sample of the plurality of training samples is derived from optical inspection of a respective training pattern; (iii) the machine learning model includes a neural network, the neural network comprising: a first series of one or more neural net layers configured to output a feature map given input data indicative of a high resolution image of a POI; a second series of neural net layers configured to output data indicative of at least one pixel block classification score given the input feature map, each pixel block classification score belonging to a region of the POI represented by pixels of a respective pixel block of the high resolution image, the pixel block classification score being indicative of a likelihood of defectivity of the respective region; thereby generating data usable to classify the POI according to a defect-related classification; The machine learning model is a) applying a first series of neural network layers to a first high-resolution image of a first POI of a first training sample of the plurality of training samples, the first POI being associated with a label indicative of defectivity, thereby generating a suspicion feature map according to a current training state of the neural network; b) applying a second series of neural network layers to the suspect feature map, thereby generating data indicative of at least one suspect pixel block score responsive to a current training state of the neural network; applying a respective suspect pixel block score to a region of the first POI represented by pixels of the respective pixel block of the first high resolution image, the respective suspect pixel block score being indicative of a likelihood of defectivity of the respective region; c) applying a first series of neural network layers to a second high resolution image of a second POI of a second training sample of the plurality of training samples, the second POI being associated with a label indicative of non-defectiveness, thereby yielding a reference feature map according to a current training state of the neural network; d) applying a second series of neural network layers to the reference feature map, thereby generating data indicative of at least one reference pixel block score responsive to a current training state of the neural network; applying each reference pixel block score to a region of the second POI represented by pixels of the respective pixel block of the second high resolution image, the reference pixel block score being indicative of a possible defect in the respective region; e) adjusting at least one weight of at least one of the first and second series of neural network layers in response to a loss function, the loss function being at least: an adjustment utilizing a distance metric derived from the suspect feature map and the reference feature map, at least one score for each suspect pixel block, and at least one score for each reference pixel block; f) repeating a) to e) for one or more additional first and second training samples of the plurality of training samples; (iv) the neural network layer is a convolutional layer; (v) the distance metric depends on the Euclidean difference between the suspect feature map and the reference feature map; (vi) the distance metric depends on the cosine similarity between the suspect feature map and the reference feature map; (vii) an additional second training sample, which is the same training sample; (viii) The loss function is further utilizing annotation data associated with the group of pixels of the first high resolution image, the annotation data indicating a defect in an area of ​​the first POI represented by the group of pixels; (ix) the annotation data is derived from human annotation of the high-resolution images; (x) comparing, by the processor, each of the at least one pixel block classification score to a defectness threshold, thereby generating an indication of whether the POI is defective; (xi) alerting an operator, by the processor, in response to an indication of whether the POI is defective; (xii) determining, by a processor, a defect bounding box in response to the at least one per-pixel block classification score; (xiii) the system a low-resolution testing tool configured to capture a low-resolution image of the POI and classify the POI according to a defectivity-related classification utilizing optical inspection of the low-resolution image; and a high resolution testing tool configured to capture a high resolution image of the POI in response to a classification of the POI as defective by the low resolution tool.

[0007] According to another aspect of the subject matter disclosed herein, there is provided a method for classifying a pattern of interest (POI) on a semiconductor specimen from a high resolution image of the pattern, the method comprising: receiving, by a processor, data indicative of a high resolution image of a POI on a specimen; and generating, by the processor, data usable to classify the POI according to a defectivity-associated classification, wherein the generating utilizes a machine learning model trained according to at least a plurality of training samples, each training sample comprising: the training patterns are similar to the POIs; Each training pattern is associated with a label that is derived from a low-resolution inspection of the respective training pattern.

[0008] This aspect of the disclosed subject matter may optionally include one or more of features (i)-(xii) listed above for the system, mutatis mutandis, in any desired combination or permutation that is technically possible.

[0009] According to another aspect of the subject matter disclosed herein, a non-transitory program storage device readable by a processing and memory circuit is provided, the non-transitory program storage device tangibly embodying computer readable instructions executable by the processing and memory circuit to perform a method of classifying a pattern of interest (POI) on a semiconductor specimen from a high resolution image of the pattern, the method comprising: receiving data indicative of a high resolution image of a POI on a specimen; and generating data usable to classify the POI according to a defectivity-related classification; The generation utilizes a machine learning model trained according to at least a plurality of training samples, each of which comprises: the training patterns are similar to the POIs; Each training pattern is associated with a label that is derived from a low-resolution inspection of the respective training pattern.

[0010] This aspect of the disclosed subject matter may optionally include one or more of features (i)-(xii) listed above for the system, mutatis mutandis, in any desired combination or permutation that is technically possible.

[0011] According to another aspect of the subject matter disclosed herein, there is provided a non-transitory program storage device readable by a processing and memory circuit, the non-transitory program storage device tangibly embodying computer readable instructions executable by the processing and memory circuit to perform a method of training a neural network to generate data indicative of at least one per-pixel block classification score given input data indicative of a high resolution image of a POI, each per-pixel block classification score pertaining to a region of the POI represented by pixels of a respective pixel block of the high resolution image, the per-pixel block classification score indicative of a likelihood of defectivity of the respective region, thereby yielding data usable for classifying the POI according to a defectivity-related classification, the training utilising at least a plurality of training samples, each training sample comprising: the training patterns are similar to the POIs; Each training pattern is associated with a label derived from a low-resolution inspection of the respective training pattern; This method is a) applying a first series of neural network layers to a first high-resolution image of a first POI of a first training sample of the plurality of training samples, the first POI being associated with a label indicative of defectivity, thereby generating a suspicion feature map according to a current training state of the neural network; b) applying a second series of neural network layers to the suspect feature map, thereby generating data indicative of at least one suspect pixel block score responsive to a current training state of the neural network; applying a respective suspect pixel block score to a region of the first POI represented by pixels of the respective pixel block of the first high resolution image, the respective suspect pixel block score being indicative of a likelihood of defectivity of the respective region; c) applying a first series of neural network layers to a second high resolution image of a second POI of a second training sample of the plurality of training samples, the second POI being associated with a label indicative of non-defectiveness, thereby yielding a reference feature map according to a current training state of the neural network; d) applying a second series of neural network layers to the reference feature map, thereby generating data indicative of at least one reference pixel block score responsive to a current training state of the neural network; applying each reference pixel block score to a region of the second POI represented by pixels of the respective pixel block of the second high resolution image, the reference pixel block score being indicative of a possible defect in the respective region; e) adjusting at least one weight of at least one of the first and second series of neural network layers in response to a loss function, the loss function being at least: an adjustment utilizing a distance metric derived from the suspect feature map and the reference feature map, at least one score for each suspect pixel block, and at least one score for each reference pixel block; and f) repeating a) through e) for one or more additional first and second training samples of the plurality of training samples.

[0012] This aspect of the disclosed subject matter may optionally include one or more of features (i)-(ix) listed above for the system, mutatis mutandis, in any desired combination or permutation that is technically possible.

[0013] In order to understand the invention and to see how it can be carried out in practice, embodiments will now be described, by way of non-limiting example, with reference to the accompanying drawings, in which: FIG. [Brief description of the drawings]

[0014] [Figure 1] 1 is a schematic block diagram of a test system in accordance with certain embodiments of the subject matter disclosed herein. [Diagram 2] 1 is a flow diagram illustrating an exemplary method for classifying patterns of interest (POI) on a semiconductor specimen according to weak labels derived from optical inspection according to certain embodiments of the subject matter disclosed herein. [Diagram 3] FIG. 1 illustrates example layers of a machine learning model that can be used to receive data indicative of an SEM image (or other high resolution image) of a POI in accordance with certain embodiments of the subject matter disclosed herein and generate data indicative of a score per pixel block that can be used to classify the POI by a defectivity-related classification. [Figure 4] FIG. 1 illustrates an exemplary method for training a machine learning model such that it can receive an input SEM image of a POI and then generate data indicative of a pixel block score that can be used to classify the POI in accordance with some embodiments of the subject matter disclosed herein. [Diagram 5] 1 is an exemplary machine learning model and training data flow according to some embodiments of the subject matter disclosed herein. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be understood by those skilled in the art that the subject matter disclosed herein may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits are not described in detail so as not to obscure the subject matter disclosed herein.

[0016] Unless specifically stated otherwise, as will be evident from the following discussion, throughout this specification, discussions utilizing terms such as "train", "obtain", "generate", "compute", "utilize", "supply", "provide", "superimpose", "apply", "adjust" and the like will be understood to refer to the operations and / or processes of a processor that manipulates and / or transforms data into other data, said data being represented as physical quantities such as electronic quantities, and / or said data representing physical objects. The term "processor" includes any computing or electronic unit having data processing circuitry capable of performing tasks based on instructions stored in a memory, such as a computer, server, chip, hardware processor, and the like. The term encompasses a single processor or multiple processors, which may be located in the same geographical area or at least partially in different areas, and capable of communicating together.

[0017] As used herein, the terms "non-transitory memory" and "non-transitory medium" should be expansively interpreted to include any volatile or non-volatile computer memory suitable for the subject matter disclosed herein.

[0018] The term "defect" as used herein should be interpreted expansively to include any type of abnormality or undesirable feature formed on or within a specimen.

[0019] The term "sample" as used herein should be expansively interpreted to include any type of wafer, mask, and other structure, combination, and / or portion thereof used to manufacture semiconductor integrated circuits, magnetic heads, flat panel displays, and other semiconductor products.

[0020] The term "testing" as used herein should be interpreted expansively to include any kind of metrology-related operation, as well as operations related to the detection and / or classification of defects in a specimen during its manufacture. Testing is provided by using a non-destructive testing tool during or after the manufacture of the specimen to be tested. As non-limiting examples, the testing process can include run-time scanning (single or multiple scans), sampling, reviewing, measuring, classifying, and / or other operations provided on the specimen or portions thereof using the same or different inspection tools. Similarly, testing can be provided prior to the manufacture of the specimen to be tested and can include, for example, generating a test recipe and / or other setup operations. It should be noted that unless specifically stated otherwise, the term "testing" as used herein or its derivatives are not limited with respect to the resolution or size of the inspection area. As non-limiting examples, various non-destructive testing tools include scanning electron microscopes, atomic force microscopes, optical inspection tools, and the like.

[0021] Embodiments of the subject matter disclosed herein are not described with reference to any particular programming language, it will be appreciated that a variety of programming languages ​​can be used to implement the teachings of the subject matter disclosed herein, as described herein.

[0022] The present invention contemplates a computer program readable by a computer for performing one or more of the methods of the present invention. The present invention further contemplates a machine-readable memory tangibly embodying a program of instructions executable by a computer for performing one or more of the methods of the present invention.

[0023] With this in mind, attention is directed to FIG. 1, which illustrates a functional block diagram of a test system in accordance with certain embodiments of the subject matter disclosed herein. The test system 100 illustrated in FIG. 1 may be used for testing a specimen (e.g., a semiconductor specimen, such as a wafer and / or portions thereof) as part of a specimen manufacturing process. The illustrated test system 100 includes a computer-based system 103 capable of automatically determining metrology-related and / or defect-related information using images of one or more specimens. The system 103 may be operatively connected to one or more low-resolution test tools 101 and / or one or more high-resolution test tools 102 and / or other test tools. The test tools are configured to capture images of the specimens and / or to review the captured images and / or to enable or realize measurements related to the captured images. The system 103 may further be operatively connected to a data repository 109. The data repository 109 may be operatively connected to a manual annotation input device 110 from which it may receive manual annotation data.

[0024] System 103 includes a processor and memory circuit (PMC) 104. PMC 104 is configured to provide the processing necessary to operate system 103, as described in further detail in various embodiments below, and comprises a processor (not separately shown) and a memory (not separately shown). In Figure 1, PMC 104 is operatively connected to a hardware-based input interface 105 and a hardware-based output interface 106.

[0025] The processor of the PMC 104 may be configured to execute several functional modules in response to computer readable instructions embodied in a non-transitory computer readable memory contained within the PMC. Such functional modules are hereinafter referred to as being contained within the PMC. The functional modules contained within the PMC 104 may include a machine learning unit (ML unit) 112. The ML unit 112 may be configured to enable data processing using machine learning models / algorithms and output application-relevant data based on images of the sample.

[0026] The ML unit 112 may include supervised or unsupervised machine learning models (e.g., for deep neural networks (DNNs)). The machine learning models of the ML unit 112 may include layers organized according to respective DNN architectures. As non-limiting examples, these DNN layers may be organized according to convolutional neural network (CNN) architectures, recurrent neural network architectures, recurrent neural network architectures, generative adversarial network (GAN) architectures, etc. Optionally, at least some of these layers may be organized in multiple DNN sub-networks. Each layer of the machine learning model may include multiple basic computation elements (CEs), typically referred to in the art as dimensions, neurons, or nodes. In some embodiments, the machine learning model may be a neural network, and each layer is a neural network layer. In some embodiments, the machine learning model may be a convolutional neural network, and each layer is a convolutional layer.

[0027] In general, a computational element in a given layer may be connected to a CE in a previous layer and / or a CE in a next layer. Each connection between a CE in a previous layer and a CE in a next layer is associated with a weight value. A given CE may receive inputs from a CE in a previous layer through each connection, and each given connection may be associated with a weight value that may be applied to the input of the given connection. The weight value may determine the relative strength of the connections and therefore the relative influence that each input has on the output of the given CE. A given CE may be configured to compute an activation value (e.g., a weighted sum of the inputs) and further derive an output by applying an activation function to the computed activations. The activation function may be, for example, an identity function, a deterministic function (e.g., linear, sigmoid, threshold, etc.), a probability function, or other suitable function.

[0028] Through each connection, the output from a given CE can be transmitted to a CE in the next layer. Similarly, as described above, each connection at the output of a CE can be associated with a weight value that is applied to the output of this CE before it is received as an input for a CE in the next layer. In addition to weight values, thresholds (including limiting functions) can be associated with connections and CEs.

[0029] The weights and / or thresholds of the machine learning model may be initially selected prior to training, and may be further iteratively adjusted or modified during training to achieve an optimal set of weights and / or thresholds for the trained ML network. After each iteration, a difference (also referred to as a loss function) between the actual output produced by the machine learning model and the target output associated with the respective training data set may be determined. This difference may be referred to as an error value. Training may be determined to be complete when a cost or loss function indicating the error value falls below a predetermined value, or when performance improvement with additional iterations is limited. Optionally, at least some of the machine learning model sub-networks (if any) may be trained separately prior to training the entire machine learning model network.

[0030] Herein, the set of machine learning model input data used to adjust the weights / thresholds of a neural network is referred to as a training set.

[0031] The system 103 can be configured to receive input data via the input interface 105, where the input data can include data created by a test tool (and / or derivatives thereof and / or metadata associated therewith) and / or data created and / or stored in one or more data repositories 109 and / or another associated data depository. It is noted that the input data can include images (e.g., captured images, images derived from captured images, simulated images, synthetic images, etc.) and associated scalar data (e.g., metadata, manual / manual image annotations, automatic attributes, etc.). It is further noted that the image data can include data related to the layer of interest of the sample and / or one or more other layers of the sample.

[0032] Upon processing the input data (e.g., low-resolution image data and / or high-resolution image data, optionally along with other data, e.g., design data, synthesis data, inspection data, etc.), system 103 can send results (e.g., instruction-related data) to any of the test tools, store the results (e.g., defect attributes, defect classifications, etc.) in storage system 107, plot the results via GUI 108, and / or send them to an external system (e.g., a FAB's revenue management system (YMS)) via output interface 106. GUI 108 can be further configured to allow user-specified inputs related to system 103.

[0033] As a non-limiting example, one or more low-resolution testing tools 101 (e.g., optical inspection system, low-resolution SEM, etc.) can test the specimen. The low-resolution testing tool 101 can transmit the resulting low-resolution image data (which can represent a low-resolution image of the specimen) to the system 103 (either directly or via one or more intermediate systems). Alternatively or additionally, the high-resolution tool 102 can test the specimen (e.g., some of the potential defect locations selected for review based on the low-resolution images can be subsequently reviewed by a scanning electron microscope (SEM) or an atomic force microscope (AFM)). The high-resolution tool 102 can transmit the resulting high-resolution image data (which can represent a high-resolution image of the specimen) to the system 103 (either directly or via one or more intermediate systems).

[0034] It should be noted that image data may be received and processed along with associated metadata (eg, pixel size, textual description of the defect type, parameters of the image capture process, etc.).

[0035] In some embodiments, image data can be received and processed along with annotation data. As a non-limiting example, a human reviewer can select an area (e.g., an oval area marked by hand) and mark it as defective, or can be marked with a label indicating defectiveness. As described below, manual or other annotation data can be utilized during training.

[0036] Those skilled in the art will readily appreciate that the teachings of the subject matter disclosed herein are not constrained by the system shown in FIG. 1 , and that equivalent and / or modified functionality may be integrated or divided in other manners and may be implemented in any suitable combination of software, firmware and / or hardware.

[0037] Without limiting the scope of the present disclosure in any way, it should also be noted that the test tools may be embodied as various types of inspection machines, such as optical imaging machines, electron beam inspection machines, etc. In some cases, the same test tool may provide low resolution image data and high resolution image data. In some cases, at least one test tool may have metrology capabilities.

[0038] It should be noted that the test system shown in Fig. 1 can be implemented in a distributed computing environment, and the above-mentioned functional modules shown in Fig. 1 can be distributed across several local and / or remote devices and linked via a communication network. It should be further noted that in another embodiment, at least some of the test tools 101 and / or 102, data repository 109, manual annotation input 110, storage system 107, and / or GUI 108 can be located outside the test system 100 and operate in data communication with the system 103 via the input interface 105 and the output interface 106. The system 103 can be implemented as a stand-alone computer used in conjunction with a test tool. Alternatively, each function of the system can be at least partially integrated with one or more test tools.

[0039] Attention is now directed to Figure 2, which shows a flow diagram illustrating an exemplary processor-based method for classifying a pattern of interest (POI) on a semiconductor specimen from a high-resolution image of the pattern in accordance with certain embodiments of the subject matter disclosed herein.

[0040] When monitoring a semiconductor manufacturing process, it may be desirable to determine various data and metrics about the specimens being manufactured, and more particularly, about the patterns on the specimens. These data / metrics may include the defect state of the pattern (e.g., defective / non-defective), the area of ​​the defect (e.g., identification of a group of pixels in a high resolution image that represents the defective pattern), and a level of confidence regarding the accuracy of the defectivity determination.

[0041] These data can be determined, for example, by human annotation of high-resolution images, but such annotation is costly and time consuming.

[0042] In some embodiments of the subject matter disclosed herein, defectiveness of regions of a high resolution image is determined using a machine learning model trained using a training set that includes training samples with "weak labels," i.e., training samples with "weak labels" are images that have associated labels that apply to the entire image (for example), and that are derived from a low resolution classification method, such as optical inspection. In the context of this disclosure, "training" can include any suitable method of configuring a machine learning model, including training methods such as those described below with reference to FIG. 4, setting parameters of the model based on the settings of another model trained using a training set that uses samples with weak labels, etc.

[0043] Some advantages of these methods include their ability to provide accurate assessments of pixel area defectivity based on readily achievable samples and without the need for annotation data.

[0044] The system 103 (e.g., the PMC 104) can receive 200 data indicative of a high-resolution image (e.g., a scanning electron microscope (SEM) image) of a POI on a sample.

[0045] The test system 100 can be configured to enable the system 103 (e.g., PMC 104) to receive the images. In some embodiments, the low-resolution test tool 101 can be configured to capture low-resolution images of a group of one or more specimens (e.g., wafers or dies) or portions of the specimens, and in so doing, capture low-resolution images of patterns of interest on the specimens. The low-resolution test tool 101 can be further configured to classify the patterns of the specimens according to a defectivity-related classification (e.g., defective / non-defective) using optical inspection (e.g., as known in the art) of the low-resolution images. The high-resolution tool 102 can be configured to capture (e.g., using an SEM) high-resolution images of the patterns of interest (POIs) in response to the classification of the POIs as defective by the one or more low-resolution test tools 101. The high-resolution tool 102 can be further configured to provide the captured high-resolution images of the POIs to the system 103.

[0046] The computer-based system 103 (e.g., PMC 104) may then generate (210) data indicative of a score for each of at least one pixel block of the high-resolution image that can be used to classify the POI according to a defect-related classification (such as defect / non-defectiveness of a region of the POI or the entire POI).

[0047] More specifically, in some embodiments, using methods described below, e.g., with reference to Figures 3-5, the computer-based system 103 (e.g., PMC 104) can generate data usable to classify POIs according to defect-related classifications. In some embodiments, the generated data indicates one or more scores (e.g., a matrix of scores), each score derived from a pixel block of the high-resolution image and indicating the defectivity (e.g., an estimate of the defectivity) of an area of ​​the POI represented by pixels of the pixel block. Such a matrix is ​​hereinafter referred to as a "grade map."

[0048] In some embodiments, the grade map generated by the system 103 (e.g., ML model 112) has the same dimensions as the input high-resolution image of the POI, and the entries of the matrix are scalar values ​​(e.g., 0-1) that indicate the likelihood that each pixel of the input image of the POI corresponds to a defective region of the POI. In some other embodiments, the generated grade map is smaller than the image of the POI (e.g., a 512x512 matrix can be generated from a 1024x1024 image, with each matrix entry containing the scores for a respective 2x2 block of pixels), in which case the scalar values ​​of the matrix indicate the likelihood that each block of pixels (e.g., a 2x2 or 4x4 block, or a block of pixels of another dimension) of the POI image corresponds to a defective region of the POI. Note that the term "block of pixels" herein can include single pixels as well as horizontal and / or vertical groups of pixels of various dimensions.

[0049] In some embodiments, the system 103 (e.g., the ML unit 112) generates the grade map by utilizing a machine learning model trained in response to a set of training examples with "weak labels" (e.g., images with associated labels that apply to the entire image). More specifically, in some such embodiments, each training example is a high-resolution image (or data indicative of a high-resolution image) captured by scanning a training pattern similar to the POI. In some embodiments, the labels associated with the training patterns are derived from optical inspection (or other low-resolution testing tool) of the respective training patterns. For example, an optical inspection tool can inspect the patterns, compare the inspected patterns to a reference pattern, and label the inspected patterns as "defective" or "non-defective." An exemplary method of training a machine learning model is described below with reference to FIG. 4.

[0050] A wafer or die can be manufactured such that multiple instances of a pattern are repeated on the wafer or die, which is then later separated into many device instances. In this disclosure, in the context of one POI being similar to another POI, the term "similar" should be interpreted broadly to include multiple instances of a pattern on a single wafer die, as well as multiple instances of a pattern across multiple instances of the same wafer or die, etc.

[0051] In some embodiments, the system 103 (e.g., the ML unit 112) may generate a data structure other than a grade map indicating per-pixel (or per-pixel block) scores that can nevertheless be used to classify the POI according to a defect-related classification.

[0052] Optionally, the system 103 (e.g., PMC 104) can then (for example) compare each of the values ​​in the grade map (or a corresponding alternative data representation) to a defectivity threshold (e.g., 0.5 on a scale of 0 to 1) (220), thereby generating an indication of whether the POI is defective and classifying the POI according to a defectivity-related classification (e.g., defective / not defective).

[0053] Optionally, the system 103 (e.g., PMC 104) may take action if there are scores in the grade map that meet the defectivity threshold. Optionally, the action may include alerting (230) an operator in response to an indication of whether a POI is defective (or that a series or multiple POIs are defective, etc.).

[0054] In some embodiments, the grade map can be used to classify the POI according to other defect-related classifications, for example, the system 103 (e.g., the PMC 104) can determine a bounding box for the defect according to the output pixel block scores.

[0055] Note that the teachings of the subject matter disclosed herein are not bound by the flowchart shown in FIG. 2. Although this flowchart has been described with reference to the elements of the system of FIG. 1 or FIG. 3, this is not binding in any way, and it should also be noted that operations can be performed by elements other than those described herein.

[0056] Next, attention is drawn to FIG. 3. FIG. 3 shows an exemplary layer of a machine learning model that can be used to receive data representing a SEM image (or other high-resolution image) of a POI, for example by PMC104 (more specifically, for example by the ML unit 112) according to a particular embodiment of the subject matter disclosed herein, and to generate data indicating scores that can be used to classify the POI by defect-related classification for each of at least one pixel block of the image.

[0057] The machine learning model can include a first series of machine learning network layers 320 (e.g., neural network layers or convolutional neural network layers) configured to receive a SEM image 310 of a POI (or another type of high-resolution image) as an input. The first series of machine learning network layers 320 can then generate a feature map 330 based on the SEM image 310.

[0058] The feature map 330 can be an intermediate output of the machine learning network. In particular, as described above, the machine learning network can include a plurality of layers L1 to L N and the feature map 330 can be obtained as the output of layer L j where 1 < j < N (in some embodiments, the intermediate layers of layers L1 to L j can be j < N and can form a convolutional neural network). As described above, in the machine learning model, each layer L N provides an intermediate output until the final output is provided by the last layer L j and this intermediate output is fed to the next layer L j+1 . For example, the SEM image 310 has dimensions X1, Y1, Z1, where - X1 corresponds to the dimension of the SEM image 310 according to the first axis X, - Y1 corresponds to the dimension of the SEM image 310 according to the second axis Y, - Z1 corresponds to the number of values ​​associated with each pixel, assuming Z1 >= 1. For example, if a representation using three colors (red, green and blue) is used, then Z1 = 3. This is not limiting and other representations can be used (for example, in an SEM microscope, electrons for each pixel are collected by several different collectors, each collector has a different position, each channel corresponds to a dimension, and thus Z1 is the total number of channels).

[0059] In some embodiments, the feature map 330 has dimensions X2, Y2, and Z2, and X2 <X1、Y2<Y1、およびZ2> Z1. Z2 is the layer L j The number of filters present in the

[0060] The machine learning model can include a second series of ML network layers 340 (e.g., neural network layers or convolutional neural network layers) configured to receive as input the feature map 330. The second series of ML network layers 340 can then generate a grade map 350 based on the feature map 330, e.g., consistent with the description above with reference to FIG.

[0061] It should be noted that the teachings of the subject matter disclosed herein are not bound to the machine learning model layers described with reference to Figure 3. Equivalent and / or modified functionality may be integrated or divided in other manners and may be implemented in any suitable combination of software, firmware and / or hardware and executed on any suitable device.

[0062] Attention is now directed to Figure 4, which illustrates an exemplary method for training a machine learning model (e.g., including a neural network having the structure described above with reference to Figure 3) such that the machine learning model can receive an input high resolution (e.g., SEM) image of a POI and then generate data indicative of a per-pixel block classification score for each of at least one pixel block of the image, the per-pixel block classification score indicative of the likelihood of defectivity of a region of the POI represented by the pixels of the pixel block. According to some embodiments of the subject matter disclosed herein, the per-pixel block classification score can be used to classify the POI according to a defectivity-related classification. This description of the method illustrated in Figure 4 makes reference to the training data flow illustrated in Figure 5.

[0063] In some embodiments, PMC 104 (e.g., ML unit 112) trains a machine learning model using pairs of training samples with weak labels (e.g., image-level labels) from a training set, where one training sample (called a suspect training sample) includes a high-resolution image (e.g., an SEM scan) of a POI that was previously labeled as defective (or suspected to be defective) and a second sample (called a reference training sample) includes a high-resolution image (e.g., an SEM scan) of a reference POI, i.e., a POI that was previously labeled (e.g., by previous classification) as non-defective.

[0064] The labeling of the high-resolution images may, for example, be derived from a low-resolution (e.g., optical) inspection of each POI. Alternatively, the labeling of the high-resolution images may be derived from human inspection or another suitable method.

[0065] In some embodiments, the training set consists entirely of training samples, where image-level labels are derived from optical inspection or other low-resolution testing tools, hi some embodiments, the training set consists of such training samples as well as other training samples.

[0066] In some embodiments, some or all of the image-level labeled training samples can be associated with annotation data, which can be derived, for example, from human annotations (e.g., a human marks an oval around a region of a high-resolution image).

[0067] The annotation data may include data indicative of the defectivity (and in some embodiments, the type of defectivity) of a portion of the specimen pattern represented by the annotated pixels of the high resolution image. In this context, the term "pixel group" may refer to a single pixel or the entire image. In this context, for example, a portion of a specimen may be considered to be represented by a unique pixel group if the portion is substantially or completely depicted by pixels.

[0068] As described below, in some embodiments, the training method illustrated in FIG. 4 adjusts weights of machine learning model layers according to a loss function computed from a distance metric derived from a suspicion feature map (i.e., a feature map derived from images of POIs labeled as defective or suspicious) and a reference feature map (i.e., a feature map derived from images of POIs labeled as non-defective).

[0069] The loss function, in some embodiments, may seek to maximize the difference between the two feature maps for feature map regions derived from pixels representing defective portions of the POI and minimize the difference between the two feature maps for feature map regions derived from pixels representing non-defective portions of the POI.

[0070] In some embodiments, the loss function thereby constitutes an attention mechanism. A first series of machine learning model layers can generate feature maps capable of identifying semantic regions, and a second series of machine learning model layers can score the regions for likelihood of defectivity.

[0071] The PMC 104 (e.g., the ML unit 112) may apply 400 a first series of neural network layers 320 to a high-resolution image 510 of a POI associated with a label indicating a defectiveness (e.g., determined or suspected defectiveness). In some embodiments, an optical inspection indicated a defectiveness of the POI, and therefore the label indicating the defectiveness was associated with the image. The first series of neural network layers 320 may then generate a feature map 530 from the high-resolution image. The feature map is generated as a function of a current training state of the neural network, and thus as training progresses, the resulting feature map changes as the training of the machine learning model progresses. A feature map generated from an image associated with a label indicating a defectiveness is referred to herein as a "suspect feature map."

[0072] The PMC 104 (e.g., the ML unit 112) may then apply 410 a second series of neural network layers 340 to the suspicion feature map 530. The second series of neural network layers 340 may then generate a grade map 560 (e.g., a score for each of one or more pixel blocks of the high resolution image 510, each score indicating the likelihood of defectness of a region of the first POI represented by the pixels of the pixel block). The scores output for the pixel blocks of the image may be referred to as pixel block scores, and thus the grade map may be referred to as pixel group block scores. The grade map is generated according to a current training state of the neural network, and thus as training progresses, the resulting grade map changes according to the progress of the training of the machine learning model. The grade map generated from the suspicion feature map is referred to herein as a "suspect grade map."

[0073] The PMC 104 (e.g., the ML unit 112) may then apply (420) a first series of neural network layers 320 to a high-resolution image 520 of the POI associated with a label indicating non-defectiveness. In some embodiments, this label was associated with the image because the image was determined to be defect-free (e.g., by optical inspection) or was otherwise determined or estimated to be defect-free. The first series of neural network layers 320 may then generate a feature map 540 from the high-resolution image 520 (computed according to a current training state of the machine learning model). The feature map generated from the image associated with the label indicating non-defectiveness is referred to herein as a "reference feature map."

[0074] The PMC 104 (e.g., the ML unit 112) may then apply (430) a second series of neural network layers 340 to the reference feature map 540. The second series of neural network layers 340 may then generate a reference grade map 570 (e.g., a score for each of one or more pixel blocks of the reference image 520, each score indicating the likelihood of defectness of an area of ​​the reference POI represented by the pixels of the pixel block) that is calculated according to a current training state of the machine learning model.

[0075] The PMC 104 (e.g., the ML unit 112) may then adjust the weights (e.g., at least one weight in at least one layer, or, e.g., all weights in all layers) of the first series of neural network layers 320 and the second series of neural network layers 340 in response to the loss function. For example, the PMC 104 (e.g., the ML unit 112) may compute (440) the loss function 590 and adjust the weights of the first series of neural network layers 320 and the second series of neural network layers 340, e.g., using backpropagation.

[0076] In some embodiments, the loss function 590 utilizes at least a distance metric (e.g., a value or set of values ​​representing the difference between the reference feature map 540 and the suspect feature map 530), the suspicion grade map 560, and the reference grade map 570. In some embodiments, the distance metric may be a difference feature map based, for example, on Euclidean distance or cosine similarity, as described below.

[0077] In some embodiments, the difference feature map 550 may be responsive to the Euclidean distance between the reference feature map 540 and the suspect feature map 530. For example, the difference feature map 550 may be computed by calculating the Euclidean distance between the reference feature map 540 and the suspect feature map 530 (i.e., in this case, the difference feature map 550 is a matrix where each entry is the arithmetic difference between corresponding entries in the two feature maps).

[0078] In other embodiments, difference feature map 550 may be a function of the cosine similarity between values ​​in reference feature map 540 and corresponding values ​​in suspect feature map 530. For example, difference feature map 550 may be a matrix in which each entry is computed by calculating the cosine similarity between values ​​in reference feature map 540 and corresponding values ​​in suspect feature map 530.

[0079] In other embodiments, the difference feature map 550 may be a different representation of the difference between the suspect feature map 530 and the reference feature map 540. In other embodiments, the loss function 590 may use a different distance metric to represent the difference between the suspect feature map 530 and the reference feature map 540.

[0080] Optionally, as described above, in some embodiments, annotation data 580 may be available. The annotation data may include data indicative of distinct pixels in suspect image 510 that have correspondence to defect regions of respective POIs. In such embodiments, loss function 590 may utilize annotation data 580 along with difference feature map 550, reference grade map 570, and suspicion grade map 560.

[0081] The PMC 104 (e.g., the ML unit 112) may repeat the application 450 of the two series of neural network layers 320, 340 to additional pairs of training samples of the plurality of training samples, and for each pair of samples, may adjust the weights of the neural network layers according to a loss function that utilizes a distance metric, such as a difference feature map 550, a suspicion grade map 560, and a reference grade map 570. For example, the PMC 104 (e.g., the ML unit 112) may train using all available suspect training samples, and along with each suspect training sample, the PMC 104 (e.g., the ML unit 112) may utilize a reference training sample from the set of training samples. In some embodiments, the PMC 104 (e.g., the ML unit 112) utilizes the same reference training sample in each iteration of training.

[0082] It should be noted that the teachings of the subject matter disclosed herein are not bound by the flow diagram shown in Figure 4, and the actions shown may occur in other orders than those shown. For example, actions 400 and 420, which are shown as successive, may be performed substantially simultaneously or in the reverse order. It should also be noted that although the flow diagram has been described with reference to the system of Figure 1, the neural network layers of Figure 3, and the data flow elements of Figure 5, this is not binding in any way, and these actions may be performed by elements other than those described herein.

[0083] It should be noted that the teachings of the subject matter disclosed herein are not constrained by the data flows described with reference to Figure 5. Equivalent and / or modified functions may be integrated or divided in other ways and may be implemented in any suitable combination of software, firmware and / or hardware and executed on any suitable device.

[0084] As described above, training a machine learning model in this manner (e.g., using suspect and reference images and utilizing an attention mechanism based on differential feature maps) can enable fast training and provide high classification accuracy.

[0085] The weight updates can use techniques such as feedforward / backpropagation methods and can rely on any optimizer (e.g., Stochastic Gradient Descent (SGD), ADAM, etc.).

[0086] It should be noted that the reference images 520 are images of reference areas of a sample (e.g., a die, a cell, etc.), with each image data representing a defect-free reference area. The reference images may be images captured from a reference (gold) die, a reference cell, or other area verified to be defect-free. Alternatively or additionally, the reference images may be simulated using CAD data and / or may be augmented after capture to exclude defects (if any) in the reference area.

[0087] It is further noted that in some embodiments, the suspect image 510 is comparable to the reference image 520 (e.g., die-to-die, cell-to-cell, die-to-database, etc.) and shows a first region of a semiconductor sample. It is envisioned that the first image shows a plurality of defects associated with the first region. The first image may be an image captured from the first region. Optionally, the first image may be further enhanced and / or may include synthetic defects introduced after capture. The first region is configured to satisfy a similarity criterion with respect to the reference region and may belong to the same or a different semiconductor sample. The similarity criterion may define, for example, that the first region and the reference region correspond to the same physical component or similar sections of the semiconductor sample (e.g., similar die, cells, etc.).

[0088] It should be noted that in some embodiments, to ensure compatibility between the images, at least one reference image 520 and the first image 510 in the training sample must undergo a registration procedure. It should be further noted that at least some of the different training samples may contain the same reference image.

[0089] It should be noted that the various features described in the various embodiments can be combined in all possible technical combinations. It should be understood that the invention is not limited in its application to the details set forth in the description contained herein or shown in the drawings. The invention is capable of other embodiments and can be practiced and carried out in various ways. It should therefore be understood that the terms and terminology used herein are for the purpose of description and should not be regarded as limiting. Thus, those skilled in the art will appreciate that the concepts on which the present disclosure is based may readily be utilized as a basis for the designing of other structures, methods and systems for carrying out some of the purposes of the subject matter disclosed herein. Those skilled in the art will readily appreciate that various modifications and changes can be made to the embodiments of the invention described above without departing from the scope of the invention as defined in the appended claims. [Explanation of symbols]

[0090] 100 Test Systems 101 Low Resolution Test Tools 102 High Resolution Test Tool 103 Computer-Based Systems, Systems 104 Processor and Memory Circuit (PMC) 105 Input Interface 106 Output Interface 107 Storage Systems 108 GUI 109 Data Repository 110 Manual Annotation Input Device 112 Machine Learning Unit (ML Unit) 310 SEM images 320 First set of machine learning network layers 330 Feature Maps 340 Second set of ML network layers 350 Grade Map 510 High Resolution Image, Suspect Image, First Image 520 High resolution image, reference image 530 feature map, suspicious feature map 540 feature maps, reference feature maps 550 Differential Feature Map 560 Grade Map, Suspect Grade Map 570 Standard Grade Map 580 Annotation Data 590 Loss Function

Claims

1. 1. A system for classifying a pattern of interest (POI) on a semiconductor specimen, comprising: a processor and memory circuit (PMC), said PMC comprising: acquiring data indicative of a high resolution image of the POI on the sample; configured to generate data usable to classify the POI according to a defect-related classification; The generating utilizes a machine learning model trained on at least a plurality of training samples, each training sample comprising: capturing high resolution training images by scanning each training pattern on a specimen with a high resolution test tool, the training pattern being similar to the POI; The high-resolution training images are obtained by associating with the labels derived from a low-resolution inspection of the respective training patterns.

2. The system of claim 1 , wherein the high resolution test tool is a scanning electron microscope (SEM) image.

3. The system of claim 1 , wherein the respective label associated with each training sample of the plurality of training samples is derived from optical inspection of the respective training pattern.

4. one or more low resolution testing tools configured to capture low resolution images of the POI and classify the POI according to the defectivity-related classification using optical inspection of the low resolution images; one or more high resolution testing tools configured to capture the high resolution image of the POI in response to a classification of the POI as defective by the one or more low resolution testing tools; The system of claim 1 further comprising:

5. The machine learning model includes a neural network, the neural network comprising: a first series of one or more neural network layers configured to output a feature map given input data indicative of the high resolution image of the POI; a second series of one or more neural network layers configured to generate data indicative of at least one per pixel block classification score given an input of the feature map; each pixel block classification score belongs to a region of the POI represented by pixels of a respective pixel block of the high resolution image, the pixel block classification score indicating a possible defect of each of the regions; The system of claim 1 , further comprising: generating data usable to classify the POI according to the defect-related classification.

6. The machine learning model, a) applying the first series of one or more neural network layers to a first high-resolution training image of a first POI of a first training sample of the plurality of training samples, the first POI being associated with a label indicative of a defect, and generating a suspicion feature map according to a current training state of the neural network; b) applying the second series of one or more neural network layers to the suspect feature map to generate data indicative of at least one per-suspect pixel block classification score responsive to the current training state of the neural network; applying a classification score for each suspect pixel block to a region of the first POI represented by pixels of a respective pixel block of the first high resolution training image, the classification score for each suspect pixel block being indicative of a likelihood of defectivity of the respective region; c) applying the first series of one or more neural network layers to a second high-resolution training image of a second POI of a second training sample of the plurality of training samples, the second POI being associated with a label indicative of non-defectiveness, and generating a reference feature map responsive to the current training state of the neural network; a) applying the second series of one or more neural network layers to the reference feature map to generate data indicative of at least one reference per-pixel block score responsive to the current training state of the neural network; d) applying a reference pixel block classification score to a region of the second POI represented by pixels of a respective pixel block of the second high resolution training image, the reference pixel block classification score being indicative of a possible defectivity of each of the regions; e) adjusting at least one weight of at least one layer of said first and second series of one or more neural network layers in response to a loss function, said loss function being at least: a distance metric derived from the suspect feature map and the reference feature map; a classification score for each of the at least one suspicious pixel blocks; and adjusting the at least one reference pixel block classification score in response to the at least one reference pixel block classification score; and f) repeating a) through e) for one or more additional first and second training samples of the plurality of training samples.

7. 6. The system of claim 5, wherein the first and second series of one or more neural network layers are convolutional layers.

8. The system of claim 5 , wherein the PMC is further configured to compare each of the at least one per-pixel block classification score to a defectness threshold and provide an indication of whether the POI is defective.

9. The system of claim 8 , wherein the PMC is further configured to alert an operator in response to the indication of whether the POI is defective.

10. 1. A method for classifying a feature of interest (POI) on a semiconductor specimen from a high resolution image of said pattern, the method being performed by a processor and memory circuit (PMC), the method comprising: receiving data indicative of a high resolution image of a POI on the sample; generating data usable to classify the POI according to a defect-related classification; The generating utilizes a machine learning model trained with at least a plurality of training samples, each training sample comprising: capturing high resolution training images by scanning each training pattern on a specimen with a high resolution test tool, the training pattern being similar to the POI; the respective training patterns are obtained by associating the respective training patterns with a label derived from a low-resolution inspection of the respective training patterns.

11. The machine learning model includes a neural network, the neural network comprising: a first series of one or more neural network layers configured to output a feature map given input data indicative of the high resolution image of the POI; a second series of neural network layers configured to generate data indicative of at least one per-pixel block classification score given the input feature map; each pixel block classification score belongs to a region of the POI represented by pixels of a respective pixel block of the high resolution image, the pixel block classification score indicating a possible defect of each of the regions; The method of claim 10 , further comprising generating data usable to classify the POI according to the defect-related classification.

12. The machine learning model, a) applying the first set of neural network layers to a first high-resolution training image of a first POI of a first training sample of the plurality of training samples, the first POI being associated with a label indicative of a defect, and generating a suspicion feature map according to a current training state of the neural network; b) applying the second series of neural network layers to the suspect feature map, resulting in data indicative of at least one per-suspect pixel block classification score responsive to the current training state of the neural network; applying a classification score for each suspect pixel block to a region of the first POI represented by pixels of a respective pixel block of the first high resolution training image, the classification score for each suspect pixel block being indicative of a likelihood of defectivity of the respective region; c) applying the first series of neural network layers to a second high-resolution training image of a second POI of a second training sample of the plurality of training samples, the second POI being associated with a label indicative of non-defectiveness, and generating a reference feature map responsive to the current training state of the neural network; d) applying the second series of neural network layers to the reference feature map to generate data indicative of at least one reference per-pixel block classification score responsive to the current training state of the neural network; applying a reference pixel block classification score to a region of the second POI represented by pixels of a respective pixel block of the second high resolution training image, the reference pixel block classification score being indicative of a likelihood of defectivity of each of the regions; e) adjusting at least one weight of at least one of the first and second series of neural network layers in response to a loss function, the loss function being at least: a distance metric derived from the suspect feature map and the reference feature map; a classification score for each of the at least one suspicious pixel blocks; and adjusting said at least one reference pixel block classification score; and f) repeating a) through e) for one or more additional first and second training samples of the plurality of training samples.

13. The method of claim 11 , further comprising: comparing each of the at least one pixel block per classification score to a defectness threshold to provide an indication of whether the POI is defective.

14. The method of claim 13 , further comprising: alerting an operator in response to the indication of whether the POI is defective.

15. The method of claim 11 , further comprising determining a defect bounding box as a function of the at least one per-pixel block classification score.

16. A computer program stored on a non-transitory computer readable storage medium comprising program instructions, the program instructions comprising: When read by a processor, the method causes the processor to execute a method for training a machine learning model to generate data usable for classifying the POI according to a defect-related classification, given input data indicative of a high resolution image of the POI, the method comprising: a) obtaining a plurality of training samples, each training sample comprising: capturing high resolution training images by scanning each training pattern on a specimen with a high resolution test tool, the training pattern being similar to the POI; obtained by associating each said training pattern with a label derived from a low-resolution inspection of said each training pattern; b) training the machine learning model with the plurality of training samples.

17. The machine learning model includes a neural network, the neural network comprising: a first series of one or more neural network layers configured to output a feature map given input data indicative of the high resolution image of the POI; a second series of neural network layers configured to generate data indicative of at least one per-pixel block classification score given an input of the feature map; each pixel block classification score belongs to a region of the POI represented by pixels of a respective pixel block of the high resolution image, the pixel block classification score indicating a possible defect of each of the regions; 17. The computer program product of claim 16, further comprising: generating data usable for classifying the POI according to the defect-related classification.

18. training the machine learning model, a) applying a first series of neural network layers to a first high-resolution training image of a first POI of a first training sample of the plurality of training samples, the first POI being associated with a label indicative of a defect, and generating a suspicion feature map according to a current training state of the neural network; b) applying a second series of neural network layers to the suspect feature map to generate data indicative of at least one classification score for each suspect pixel block in response to the current training state of the neural network; applying a classification score for each suspect pixel block to a region of the first POI represented by pixels of a respective pixel block of the first high resolution training image, the classification score for each suspect pixel block being indicative of a likelihood of defectivity of the respective region; c) applying the first series of neural network layers to a second high-resolution training image of a second POI of a second training sample of the plurality of training samples, the second POI being associated with a label indicative of non-defectiveness, and generating a reference feature map according to the current training state of the neural network; d) applying the second series of neural network layers to the reference feature map to generate data indicative of at least one reference per-pixel block classification score responsive to the current training state of the neural network; applying a reference pixel block classification score to a region of the second POI represented by pixels of a respective pixel block of the second high resolution image, the reference pixel block classification score being indicative of a possible defectivity of each of the regions; e) adjusting at least one weight of at least one of the first and second series of neural network layers in response to a loss function, the loss function being at least: a distance metric derived from the suspect feature map and the reference feature map; a classification score for each of the at least one suspicious pixel blocks; and adjusting said at least one reference pixel block classification score; and f) repeating a) through e) for one or more additional first and second training samples of the plurality of training samples.

19. 20. The computer program product of claim 18, wherein the distance metric is a difference map calculated according to a Euclidean difference between the suspect feature map and the reference feature map, or a difference map calculated according to a cosine similarity between the suspect feature map and the reference feature map.

20. 20. The computer program product of claim 18, wherein the loss function further utilizes annotation data associated with pixels of the first high resolution image, the annotation data indicating a defectivity of a region of the first POI represented by the pixels.

Images