Substrate inspection method, substrate inspection program, and substrate inspection device

Abstract

This substrate inspection method includes acquiring a reference input image based on a captured image of a surface of a substrate for reference, acquiring reference intermediate information generated by an intermediate layer of a neural network, constructed in advance so as to output a result of recognition of an inputted image, when the reference input image is inputted to the neural network, generating a reference feature image indicating a feature of the reference input image on the basis of the reference intermediate information, acquiring an inspection input image based on a captured image of the surface of a substrate to be inspected, acquiring inspection intermediate information generated by an intermediate layer of the neural network when the inspection input image is inputted to the neural network, generating an inspection feature image indicating a feature of the inspection input image on the basis of the inspection intermediate information, and determining whether an abnormality is present in the surface of the substrate to be inspected on the basis of the result of comparing the reference feature image and the inspection feature image.

Description

[Technical field]

[0001] The present disclosure relates to a substrate inspection method, a substrate inspection program, and a substrate inspection apparatus. [Background technology]

[0002] Patent Document 1 discloses an apparatus that classifies defects occurring on a substrate based on a captured image, which is an inspection target obtained by capturing an image of the substrate. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2019-124591 A Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure provides a substrate inspection method, a substrate inspection program, and a substrate inspection apparatus that are useful for accurately detecting abnormalities on a substrate surface. [Means for solving the problem]

[0005] A substrate inspection method according to one aspect of the present disclosure includes obtaining a reference input image based on an image captured of a surface of a reference substrate, obtaining reference intermediate information generated in an intermediate layer of a neural network when the reference input image is input into a neural network pre-constructed to output a recognition result of the input image, generating a reference feature image indicating characteristics of the reference input image based on the reference intermediate information, obtaining an inspection input image based on an image captured of the surface of a substrate to be inspected, obtaining inspection intermediate information generated in an intermediate layer of the neural network when the inspection input image is input into the neural network, generating an inspection feature image indicating characteristics of the inspection input image based on the inspection intermediate information, and determining the presence or absence of an abnormality on the surface of the substrate to be inspected based on a result of comparing the reference feature image and the inspection feature image. Effect of the Invention

[0006] According to the present disclosure, there are provided a substrate inspection method, a substrate inspection program, and a substrate inspection apparatus that are useful for detecting abnormalities on a substrate surface with high accuracy. [Brief description of the drawings]

[0007] [Figure 1] FIG. 1 is a perspective view diagrammatically illustrating a substrate processing system according to a first embodiment. [Diagram 2] FIG. 2 is a side view diagrammatically illustrating an example of a coating and developing apparatus. [Diagram 3] FIG. 3 is a schematic diagram showing an example of the inspection unit. [Figure 4] FIG. 4 is a block diagram illustrating an example of a functional configuration of the control device. [Diagram 5] FIG. 5 is a block diagram illustrating an example of a hardware configuration of the control device. [Figure 6] FIG. 6 is a schematic diagram showing an example of a substrate inspection method. [Figure 7] FIG. 7 is a flowchart showing an example of an inspection process in the preparation phase. [Figure 8]FIG. 8 is a diagram illustrating an example of an extracted image extracted from the intermediate layer. [Figure 9] FIG. 9 is a diagram for explaining an example of a calculation process of the Mahalanobis distance. [Figure 10] FIG. 10 is a graph for conceptually explaining an example of the Mahalanobis distance. [Figure 11] FIG. 11 is a flowchart showing an example of an inspection process in the production phase. [Figure 12] FIG. 12 is a schematic diagram showing an example of a substrate inspection method according to the second embodiment. [Figure 13] FIG. 13 is a schematic diagram showing an example of a substrate inspection method according to the second embodiment. [Figure 14] FIG. 14 is a flowchart showing an example of a series of processes executed in the substrate inspection method. [Figure 15] FIG. 15 is a flowchart showing an example of the test preparation process. [Figure 16] FIG. 16 is a flowchart showing an example of an inspection process for a workpiece. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Hereinafter, some embodiments will be described with reference to the drawings. In the description, the same elements or elements having the same functions are given the same reference numerals, and duplicated descriptions will be omitted. Some drawings show an orthogonal coordinate system defined by an X-axis, a Y-axis, and a Z-axis. In the following description, the Z-axis corresponds to the up-down direction, and the X-axis and the Y-axis correspond to the horizontal direction.

[0009] [First embodiment] First, a substrate processing system according to a first embodiment will be described with reference to FIGS. 1 to 11. The substrate processing system 1 shown in FIG. 1 is a system that performs the formation of a photosensitive film on a workpiece W, the exposure of the photosensitive film, and the development of the photosensitive film. The workpiece W to be processed is, for example, a substrate, or a substrate on which a film, a circuit, or the like is formed by performing a predetermined process. The substrate included in the workpiece W is, for example, a wafer containing silicon. The workpiece W (substrate) may be formed in a circular shape. The workpiece W to be processed may be a glass substrate, a mask substrate, an FPD (Flat Panel Display), or the like, or may be an intermediate body obtained by performing a predetermined process on such a substrate. The photosensitive film is, for example, a resist film.

[0010] The substrate processing system 1 includes a coating and developing apparatus 2 and an exposure apparatus 3. The exposure apparatus 3 performs an exposure process on a resist film (photosensitive coating) formed on a workpiece W (substrate). Specifically, the exposure apparatus 3 irradiates an exposure target portion of the resist film with energy rays by a method such as immersion exposure. The coating and developing apparatus 2 performs a process of forming a resist film on the surface of the workpiece W before the exposure process by the exposure apparatus 3, and performs a development process of the resist film after the exposure process.

[0011] [Substrate processing equipment] 1 and 2, the coating and developing apparatus 2 includes a carrier block 4, a processing block 5, an interface block 6, and a control device 100.

[0012] The carrier block 4 introduces the workpiece W into the coating and developing apparatus 2 and removes the workpiece W from the coating and developing apparatus 2. For example, the carrier block 4 can support a plurality of carriers C (containers) for the workpieces W, and has a built-in transport device A1 including a delivery arm. The carrier C stores a plurality of circular workpieces W, for example. The transport device A1 removes the workpiece W from the carrier C and delivers it to the processing block 5, and receives the workpiece W from the processing block 5 and returns it to the carrier C. The processing block 5 has a plurality of processing modules 11, 12, 13, and 14.

[0013] The processing module 11 incorporates a liquid processing unit U1, a heat processing unit U2, an inspection unit U3, and a transport device A3 that transports the workpiece W to these units. The processing module 11 forms an underlayer film on the surface of the workpiece W using the liquid processing unit U1 and the heat processing unit U2. The liquid processing unit U1 of the processing module 11 applies a processing liquid for forming the underlayer film onto the workpiece W. The heat processing unit U2 of the processing module 11 performs various heat treatments associated with the formation of the underlayer film. The inspection unit U3 performs processing to inspect the condition of the surface of the workpiece W before the formation of the underlayer film, after the formation of the underlayer film, or before the processing liquid for forming the underlayer film is applied and heat treatment is performed.

[0014] The processing module 12 incorporates a liquid processing unit U1, a heat processing unit U2, an inspection unit U3, and a transport device A3 that transports the workpiece W to these units. The processing module 12 forms a resist film on the underlayer film using the liquid processing unit U1 and the heat processing unit U2. The liquid processing unit U1 of the processing module 12 applies a processing liquid (resist) for forming a resist film onto the underlayer film. The heat processing unit U2 of the processing module 12 performs various heat treatments associated with the formation of the resist film. The inspection unit U3 performs processing to inspect the condition of the surface of the workpiece W before the resist film is formed, after the resist film is formed, or before the resist is applied and heat treatment is performed.

[0015] The processing module 13 incorporates a liquid processing unit U1, a heat processing unit U2, an inspection unit U3, and a transport device A3 that transports the workpiece W to these units. The processing module 13 forms an upper layer film on the resist film using the liquid processing unit U1 and the heat processing unit U2. The liquid processing unit U1 of the processing module 13 applies a processing liquid for forming the upper layer film onto the resist film. The heat processing unit U2 of the processing module 13 performs various heat treatments associated with the formation of the upper layer film. The inspection unit U3 performs processing to inspect the condition of the surface of the workpiece W before the upper layer film is formed, after the upper layer film is formed, or before the processing liquid for forming the upper layer film is applied and heat treatment is performed.

[0016] The processing module 14 incorporates a liquid processing unit U1, a heat processing unit U2, an inspection unit U3, and a transport device A3 that transports the workpiece W to these units. The processing module 14 performs a developing process on the resist film after exposure using the liquid processing unit U1 and the heat processing unit U2. The liquid processing unit U1 of the processing module 14 performs the developing process on the resist film, for example, by supplying a developing solution onto the surface of the exposed workpiece W and then rinsing it away with a rinsing solution.

[0017] The heat treatment unit U2 of the processing module 14 performs various heat treatments associated with the development process. Specific examples of heat treatments include a heat treatment before the development process (PEB: Post Exposure Bake), a heat treatment after the development process (PB: Post Bake), etc. The inspection unit U3 performs a process for inspecting the condition of the surface of the workpiece W before the development process and PEB are performed, after the development process and PB are performed, or before the developer is supplied and PB is performed.

[0018] A shelf unit U10 is provided on the carrier block 4 side in the processing block 5. The shelf unit U10 is divided into multiple cells arranged in the vertical direction. A transport device A7 including a lifting arm is provided near the shelf unit U10. The transport device A7 lifts and lowers the workpiece W between the cells of the shelf unit U10.

[0019] A shelf unit U11 is provided on the interface block 6 side in the processing block 5. The shelf unit U11 is partitioned into a plurality of cells arranged in the vertical direction.

[0020] The interface block 6 transfers the workpiece W to and from the exposure apparatus 3. For example, the interface block 6 has a built-in transport device A8 including a transfer arm, and is connected to the exposure apparatus 3. The transport device A8 transfers the workpiece W placed on the shelf unit U11 to the exposure apparatus 3, receives the workpiece W from the exposure apparatus 3, and returns it to the shelf unit U11.

[0021] The control device 100 controls each device included in the coating and developing apparatus 2 to perform a coating and developing process (substrate processing) in the following manner, for example: First, the control device 100 controls the transport device A1 to transport the workpiece W in the carrier C to the shelf unit U10, and controls the transport device A7 to place the workpiece W in a cell for the processing module 11.

[0022] Next, the control device 100 controls the transport device A3 to transport the workpiece W from the shelf unit U10 to the liquid processing unit U1 in the processing module 11. The control device 100 controls the liquid processing unit U1 to form a film of the processing liquid for forming the underlayer film on the surface of the workpiece W. The control device 100 controls the heat processing unit U2 to heat the workpiece W on which the film of the processing liquid for forming the underlayer film has been formed to form the underlayer film. Thereafter, the control device 100 controls the transport device A3 to return the workpiece W on which the underlayer film has been formed to the shelf unit U10, and controls the transport device A7 to place the workpiece W in a cell for the processing module 12. The control device 100 may control the inspection unit U3 to inspect the surface of the workpiece W at any timing during the processing in the processing module 11.

[0023] Next, the control device 100 controls the transport device A3 to transport the workpiece W from the shelf unit U10 to the liquid processing unit U1 in the processing module 12. The control device 100 controls the liquid processing unit U1 to form a film of the processing liquid for forming a resist film on the surface of the workpiece W. The control device 100 controls the heat processing unit U2 to heat the workpiece W on which the film of the processing liquid for forming a resist film has been formed, to form a resist film. Thereafter, the control device 100 controls the transport device A3 to return the workpiece W to the shelf unit U10, and controls the transport device A7 to place the workpiece W in a cell for the processing module 13. The control device 100 may control the inspection unit U3 to inspect the surface of the workpiece W at any timing during the processing in the processing module 12.

[0024] Next, the control device 100 controls the transport device A3 to transport the workpiece W on the shelf unit U10 to the liquid processing unit U1 in the processing module 13. The control device 100 also controls the liquid processing unit U1 to form a film of the processing liquid for forming an upper layer film on the resist film of the workpiece W. The control device 100 controls the heat processing unit U2 to heat the workpiece W on which the film of the processing liquid for forming an upper layer film has been formed, to form an upper layer film. Thereafter, the control device 100 controls the transport device A3 to transport the workpiece W to the shelf unit U11. The control device 100 may control the inspection unit U3 to inspect the surface of the workpiece W at any timing during the processing in the processing module 13.

[0025] Next, the control device 100 controls the transport device A8 to send the workpiece W on the shelf unit U11 to the exposure device 3. Thereafter, the control device 100 controls the transport device A8 to receive the workpiece W that has been subjected to the exposure process from the exposure device 3 and place it in a cell for the processing module 14 in the shelf unit U11.

[0026] Next, the control device 100 controls the transport device A3 to transport the workpiece W on the shelf unit U11 to each unit in the processing module 14, and controls the liquid processing unit U1 and the heat processing unit U2 to perform a developing process of the resist film on the workpiece W. Thereafter, the control device 100 controls the transport device A3 to return the workpiece W to the shelf unit U10, and controls the transport device A7 and the transport device A1 to return the workpiece W into the carrier C. The control device 100 may control the inspection unit U3 to inspect the surface of the workpiece W at any timing of the processing in the processing module 14. This completes the coating and developing process for one workpiece W. The control device 100 controls each device of the coating and developing apparatus 2 to perform the coating and developing process for each of the subsequent multiple workpieces W in the same manner as described above.

[0027] The specific configuration of the substrate processing apparatus is not limited to the above-described configuration of the coating and developing apparatus 2. The substrate processing apparatus may be any type as long as it includes a unit for inspecting the surface of the workpiece W to be subjected to a predetermined process and a control device for controlling this unit.

[0028] (Inspection unit) Next, the inspection unit U3 included in the processing modules 11 to 14 will be described. The inspection unit U3 has a function of capturing an image of the surface of the workpiece W (hereinafter referred to as "surface Wa") to acquire image data. The inspection unit U3 may capture an image of the entire surface Wa of the workpiece W to acquire image data of the entire surface Wa. As shown in FIG. 3, the inspection unit U3 includes, for example, a housing 30, a holding unit 31, a linear driving unit 32, an imaging unit 33, and a light projecting / reflecting unit 34.

[0029] The holder 31 holds the workpiece W horizontally with the surface Wa facing upward. The linear drive unit 32 includes a power source such as an electric motor, and moves the holder 31 along a horizontal linear path. The imaging unit 33 has a camera 35 such as a CCD camera. The camera 35 is provided near one end of the inspection unit U3 in the moving direction of the holder 31, and is directed toward the other end in the moving direction. The light projecting / reflecting unit 34 projects light into the imaging range, and guides the reflected light from the imaging range to the camera 35. For example, the light projecting / reflecting unit 34 has a half mirror 36 and a light source 37. The half mirror 36 is provided in a position higher than the holder 31, in the middle of the moving range of the linear drive unit 32, and reflects light from below to the camera 35. The light source 37 is provided above the half mirror 36, and irradiates illumination light downward through the half mirror 36.

[0030] The inspection unit U3 operates as follows to acquire image data of the surface Wa of the workpiece W. First, the linear drive unit 32 moves the holder 31. This causes the workpiece W to pass under the half mirror 36. During this passage process, reflected light from each part of the surface Wa of the workpiece W is sent sequentially to the camera 35. The camera 35 forms an image of the reflected light from each part of the surface Wa of the workpiece W, and acquires image data of the surface Wa of the workpiece W (the entire surface Wa). The captured image obtained by capturing the surface Wa of the workpiece W changes depending on the state of the surface Wa of the workpiece W. In other words, acquiring a captured image (captured image data) of the surface Wa of the workpiece W corresponds to acquiring information indicating the state of the surface Wa of the workpiece W.

[0031] The captured image data acquired by the camera 35 is sent to the control device 100. In the control device 100, the condition of the surface Wa of the workpiece W can be inspected based on the captured image data of the surface Wa. For example, the presence or absence of defects on the surface Wa of the workpiece W can be inspected. In the present disclosure, image data in which pixel values ​​for each pixel are defined may be simply referred to as an "image."

[0032] [Control device] 4, the control device 100 has a process control unit 102 and an inspection control unit 110 as functional components (hereinafter referred to as "functional modules"). The processes executed by the process control unit 102 and the inspection control unit 110 correspond to the processes executed by the control device 100. The process control unit 102 controls the liquid processing unit U1 and the heat processing unit U2 so as to perform the liquid processing and heat processing in the above-mentioned coating and developing process on the workpiece W.

[0033] The inspection control unit 110 (substrate inspection device) inspects the workpiece W based on image data obtained from the inspection unit U3 at any stage when performing the coating and developing process. The inspection of the workpiece W includes determining whether or not there is an abnormality (defect) on the surface Wa of the workpiece W. Defects on the surface Wa include, for example, scratches, adhesion of foreign matter, uneven application of the processing liquid, and non-application of the processing liquid.

[0034] Prior to inspection, the inspection control unit 110 prepares reference data to be used in the inspection from a reference workpiece W (reference substrate). The inspection control unit 110 performs inspection of the workpiece W to be inspected (substrate to be inspected) based on the reference data. The reference workpiece W and the workpiece W to be inspected are the same type of workpiece (substrate). The reference workpiece W and the workpiece W to be inspected are subjected to a coating and developing process under the same processing conditions, and the preparation of the reference data and the inspection of the workpiece W are performed at the same timing in the coating and developing process (for example, after the application of resist and before heat treatment).

[0035] The inspection control unit 110 has, as functional modules, a first input image acquisition unit 112, a first intermediate information acquisition unit 114, a first feature image generation unit 116, a reference image holding unit 118, a model holding unit 132, a second input image acquisition unit 122, a second intermediate information acquisition unit 124, a second feature image generation unit 126, an abnormality determination unit 136, and a determination result output unit 138. The processing executed by each functional module of the inspection control unit 110 corresponds to the processing executed by the inspection control unit 110 (control device 100).

[0036] The first input image acquisition unit 112 acquires a reference input image based on an image captured of a reference workpiece W. The first intermediate information acquisition unit 114 acquires reference intermediate information generated in an intermediate layer of a neural network (hereinafter referred to as "image recognition model M") that is constructed in advance to output a recognition result of an input image, when the reference input image is input to the neural network. The first feature image generation unit 116 generates a reference feature image that indicates the features of the reference input image, based on the reference intermediate information.

[0037] The reference image holding unit 118 holds (stores) the reference feature image generated by the first feature image generating unit 116. The reference feature image generated by the first feature image generating unit 116 is reference data used in inspecting the workpiece W to be inspected. The model holding unit 132 holds the image recognition model M.

[0038] The second input image acquisition unit 122 acquires an inspection input image based on an image captured of the surface Wa of the workpiece W to be inspected. The second intermediate information acquisition unit 124 acquires inspection intermediate information generated in the intermediate layer of the image recognition model M when the inspection input image is input to the image recognition model M. The second feature image generation unit 126 generates an inspection feature image indicating the features of the inspection input image based on the inspection intermediate information.

[0039] The abnormality determination unit 136 determines the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on the result of comparing the reference feature image with the inspection feature image. The determination result output unit 138 outputs the determination result by the abnormality determination unit 136. When the abnormality determination unit 136 determines that there is an abnormality on the surface Wa of the workpiece W, the determination result output unit 138 may output an abnormality signal indicating that the workpiece W to be inspected is abnormal. The determination result output unit 138 may output the abnormality signal to the process control unit 102, may output it to a higher-level controller, or may output it to an output device such as a monitor for notifying an operator or the like of information.

[0040] The control device 100 is configured with one or more computers. The control device 100 has, for example, a circuit 150 shown in FIG. 5. The circuit 150 has one or more processors 152, a memory 154, a storage 156, and an input / output port 158. The storage 156 has a computer-readable storage medium, such as a hard disk. The storage medium stores a program (substrate inspection program) for causing the control device 100 to execute a substrate inspection method described below. The storage medium may be a removable medium, such as a non-volatile semiconductor memory, a magnetic disk, or an optical disk.

[0041] Memory 154 temporarily stores the programs loaded from the storage medium of storage 156 and the results of calculations by processor 152. Processor 152 configures the above-mentioned functional modules by executing the above-mentioned programs in cooperation with memory 154. Input / output port 158 ​​inputs and outputs electrical signals between liquid processing unit U1, heat processing unit U2, inspection unit U3, etc., in accordance with instructions from processor 152.

[0042] The hardware configuration of the control device 100 is not necessarily limited to configuring each functional module by a program. For example, each functional module of the control device 100 may be configured by a dedicated logic circuit or an ASIC (Application Specific Integrated Circuit) that integrates the dedicated logic circuit. When the control device 100 is configured by multiple computers (multiple circuits), some of the functional modules may be realized by one computer (circuit) and the remaining part of the functional modules may be realized by another computer (circuit).

[0043] [Board inspection method] Next, a series of processes executed by the control device 100 (inspection control unit 110) will be described as an example of a substrate inspection method. For example, as shown in FIG. 6, the control device 100 executes processes in a preparation phase and in a production phase. In the preparation phase, the control device 100 applies a coating and developing process to a reference workpiece W, and then executes preparations for inspection of the workpiece W in the production phase. In the production phase, the control device 100 applies a coating and developing process to a plurality of workpieces W in sequence, and then inspects each workpiece W to which the coating and developing process is applied. The workpiece W inspected in the production phase corresponds to the workpiece W to be inspected described above.

[0044] <Preparation Phase> Fig. 7 is a flowchart showing a series of processes in the preparation phase shown in Fig. 6. The series of processes shown in Fig. 7 starts in a state where a reference workpiece W that has been subjected to a pre-inspection process in the coating and developing process and has been determined to be normal is transported to the inspection unit U3.

[0045] The control device 100 first executes step Sa-1. In step Sa-1, for example, the first input image acquisition unit 112 of the inspection control unit 110 acquires a captured image PIr of the surface Wa of the reference workpiece W by capturing an image of the surface Wa of the reference workpiece W using the inspection unit U3. The captured image PIr may be a color image. The captured image PIr may include the entire surface Wa, and the number of pixels in the horizontal direction and the number of pixels in the vertical direction in the captured image PIr may be the same.

[0046] Next, the control device 100 executes step Sa-2. In step Sa-2, for example, the first input image acquisition unit 112 generates an enhanced image EIr by performing a process for enhancing contrast on the captured image PIr obtained in step Sa-1. By performing the contrast enhancement process, for example, the difference (difference in luminance) between a bright part and a dark part on the image is enhanced. The first input image acquisition unit 112 may perform the contrast enhancement process by various methods. The first input image acquisition unit 112 may perform the contrast enhancement process on the captured image PIr by transforming (adjusting) a tone curve. The first input image acquisition unit 112 may perform the contrast enhancement process by applying a known spatial filter to the captured image PIr.

[0047] Next, the control device 100 executes step Sa-3. In step Sa-3, for example, the first intermediate information acquisition unit 114 acquires reference intermediate information generated in an intermediate layer of the image recognition model M when the enhanced image EIr (reference input image) obtained in step Sa-2 is input to the image recognition model M. The first intermediate information acquisition unit 114 acquires, as reference intermediate information, an extracted image group CIGr including a plurality of extracted images CIr (a plurality of reference extracted images) generated based on the enhanced image EIr and a plurality of filters that extract different features in the intermediate layer of the image recognition model M, for example.

[0048] Here, the image recognition model M used in step Sa-3 will be described. The image recognition model M is a model constructed in advance by machine learning so as to output a result (recognition result) of classifying the contents contained in an image into categories when the image is input. The image recognition model M may be a multi-layered neural network constructed by deep learning. The image recognition model M may be a CNN (Convolutional Neural Network).

[0049] The image recognition model M does not have to be a model constructed to classify the work W in an image into a category according to a predetermined condition. The image recognition model M may be a model that recognizes the type of object (e.g., animal, fruit), a model that recognizes a human face, or a model that recognizes characters. The CNN may be composed of an input layer, multiple convolution layers, a pooling layer, a fully connected layer, and an output layer.

[0050] In the convolution layer (intermediate layer) included in the image recognition model M, multiple filters are used, and convolution is performed on the input image to that layer. The filter is also called a kernel, and each file is lattice-shaped numerical data that represents a specific shape (characteristic). The size of the filter is smaller than the size of the input image. The multiple filters are set in the convolution layer so that different shapes (characteristics) are obtained from each other. In the convolution calculation of the input image using one file, for example, a conversion process is performed in which a product is calculated for each pixel between a partial image (window) of the same size as the filter in the input image and the filter, and the sum of the calculation results of the products of all pixels is calculated. Then, the conversion process is repeated on the entire input image while moving the position of the partial image by a predetermined number of pixels.

[0051] By repeating the conversion process, an image in which the shape set by the filter is extracted (responding to the shape) is obtained as a convolution result. The convolution result is called a feature map. The multiple extracted images CIr acquired by the first intermediate information acquisition unit 114 are multiple images obtained by performing convolution using N filters in any one of the multiple convolution layers. N is a natural number equal to or greater than 2. The first intermediate information acquisition unit 114 may input the enhanced image EIr obtained in step Sa-2 to the image recognition model M held by the model holding unit 132, and then acquire multiple extracted images CIr from intermediate calculation results by the image recognition model M.

[0052] FIG. 8 shows a schematic diagram of an extraction image group CIG obtained by inputting an image of the surface Wa of the workpiece W to the image recognition model M and performing convolution using N filters in one of the multiple convolution layers. The extraction image group CIG includes multiple extraction images CI. When the input image to the image recognition model M is the enhanced image EIr, multiple extraction images CI obtained correspond to the multiple extraction images CIr. The multiple extraction images CI include extraction images CI1, CI2, . . . , CIN. N is, for example, 230 to 270. In the following, a case in which the number of vertical pixels of one extraction image CI is 255 and the number of horizontal pixels is 255 is illustrated. The multiple extraction images CI may be grayscale images. When a color image is input, the image recognition model M may convert the image to grayscale before performing calculations. In FIG. 8, extracted images other than extracted images CI1, CI2, CI3, CIj, and CIN are shown simplified with simple circles, but these extracted images also have pixel values.

[0053] Returning to Figs. 6 and 7, after executing step Sa-3, the control device 100 executes step Sa-0. In step Sa-0, for example, the control device 100 judges whether or not the series of processes of steps Sa-1 to Sa-3 have been executed for a predetermined number of reference workpieces W. The predetermined number is set, for example, to a number that can eliminate individual differences between the reference workpieces W. If it is determined that the series of processes has not been executed for the predetermined number of reference workpieces W (step Sa-0: NO), the process executed by the control device 100 returns to step Sa-1. Then, the control device 100 executes the series of processes of steps Sa-1 to Sa-3 for different individual reference workpieces W.

[0054] In step Sa-0, if it is determined that a series of processes has been performed for a predetermined number of reference workpieces W (step Sa-0: YES), the process executed by the control device 100 proceeds to step Sa-4. In step Sa-4, for example, the first feature image generating unit 116 performs a calculation to generate a reference feature image DIr based on the multiple extraction images CIr acquired in step Sa-3. The first feature image generating unit 116 calculates a Mahalanobis distance (reference Mahalanobis distance) for array data of pixel values ​​(luminance values) for each pixel included in the multiple extraction images CIr, based on the data distribution of the multiple extraction images CIr. The first feature image generating unit 116 may calculate the Mahalanobis distance for each of the multiple reference workpieces W (for each reference workpiece W).

[0055] Here, an example of a method for calculating the Mahalanobis distance and a method for generating a reference feature image DIr will be described with reference to FIG. 9 and FIG. 10. In the following, the vertical coordinate on the image is represented by "i" and the horizontal coordinate is represented by "j". A pixel (i, j) indicates a pixel located in the i-th row and j-th column, and each of i and j is a natural number from 1 to N. First, array data of pixel values ​​is created for extracted images CIr1, CIr2, . . . , CIrN (plural extracted images) obtained from the first reference work W. The pixel values ​​of all pixels included in the extracted image CIr1 can be represented by one vertically aligned array data. The pixel values ​​included in the extracted images CIr2, . . . , CIrN can also be represented by one vertically aligned array data. The number of data in one vertically aligned array data obtained from one reference work W is, for example, 255×255=65025.

[0056] Next, array data of pixel values ​​is created for extracted images CIr1, CIr2, . . . , CIrN obtained from the second reference work W, and is arranged below the array data for the first reference work W. Similarly, array data of pixel values ​​is created for the third and subsequent reference work W, and is arranged in order below the array data already created. In this case, if the number of reference work W is A (A is a natural number greater than or equal to 2), the number of data in one vertical row of array data is, for example, 65025 x A.

[0057] In FIG. 9, pixel values ​​of all pixels in an extracted image CIr1 obtained from each of a plurality of reference workpieces W are shown as array data arranged vertically as a variable x1. Pixel values ​​of all pixels in an extracted image CIr2 obtained from each of a plurality of reference workpieces W are shown as array data arranged vertically as a variable x2. Similarly, pixel values ​​of all pixels in extracted images CIr3 to CIrN-1 obtained from each of a plurality of reference workpieces W are shown as array data arranged vertically as variables x3 to xN-1, respectively. Pixel values ​​of all pixels in an extracted image CIrN obtained from each of a plurality of reference workpieces W are shown as array data arranged vertically as a variable xN. N pieces of array data indicated by variables x1 to xN are arranged in order horizontally. The N pieces of array data indicated by variables x1 to xN include N variables, and are therefore N-dimensional data.

[0058] In each vertically arranged array data, the order of the coordinates is the same. Therefore, in multiple horizontally arranged array data, pixel values ​​of pixels (i,j) with the same coordinates in variables x1 to xN are arranged horizontally. For example, for the first reference work W, the pixel value of variable x1 in pixel (1,1), the pixel value of variable x2 in pixel (1,1), and the pixel values ​​of variables x3 to xN in pixel (1,1) are arranged in the first row of the array data in FIG. 9. Here, any variable among variables x1 to xN is represented as "xn", where n is any one of natural numbers from 1 to N. The pixel value of pixel (i,j) of a specific coordinate in variable xn is represented as "xn[i,j]". x1[i,j], x2[i,j], ..., and xN[i,j] are arranged horizontally in this order in the array data. As described above, the first feature image generating unit 116 executes the process of arranging the values ​​(pixel values) included in each of the multiple extraction images CIr in a vertical array for the multiple reference works W.

[0059] Next, the first feature image generating unit 116 calculates average array data indicating the average for each variable and a covariance matrix in order to calculate the Mahalanobis distance. FIG. 10 shows a graph for explaining the concept of the Mahalanobis distance. Here, for the sake of simplicity, the concept of the Mahalanobis distance is explained, which is obtained from one reference work W and calculated from two-dimensional array data of only variables x1 and x2. In the graph shown in FIG. 10, the horizontal axis is the variable x1, and the vertical axis is the variable x2. In the graph in FIG. 10, the value of the combination of variables x1 and x2 (x1[i,j], x2[i,j]) is plotted for each pixel (coordinate).

[0060] The value of the pixel value combination indicated by [m1, n1] and the value of the pixel value combination indicated by [m2, n2] are approximately the same in distance from the value of the combination of Mean(x1) which is the average value of variable x1 and Mean(x2) which is the average value of variable x2. m1, m2, n1, and n2 are any one of natural numbers from 1 to N. However, the value of the pixel value combination indicated by [m2, n2] is out of the distribution of the values ​​of the combination of variables x1 and x2 compared to the value of the pixel value combination indicated by [m1, n1]. The Mahalanobis distance can represent the degree of deviation (degree of anomaly) of variables x1 and x2 from the data distribution.

[0061] The first feature image generating unit 116 calculates the average array data (average) shown in FIG. 9 by calculating the average μn of pixel values ​​for each variable xn (for each of the variables x1 to xN). The average μn is calculated from one vertically aligned array data included in the variable xn, and is the arithmetic average of pixel values ​​of all pixels related to all reference works W in that column. The first feature image generating unit 116 calculates the variance σn from the vertically aligned array data of pixel values ​​for each variable xn (for each of the variables x1 to xN). The first feature image generating unit 116 calculates the correlation coefficient Srs (covariance) for each of all combinations of two variables among the variables x1 to xN. Each of r and s in the correlation coefficient Srs satisfies r≠s and is a natural number from 1 to N. A covariance matrix is ​​obtained by calculating the variance σn and the correlation coefficient Srs.

[0062] The first feature image generating unit 116 uses the average array data and the covariance matrix to calculate the Mahalanobis distance for array data of pixel values ​​arranged horizontally for each pixel (i, j) in variables x1 to xN related to one reference work W. The first feature image generating unit 116 calculates the Mahalanobis distance from the array data of pixel values ​​of each pixel for all pixels related to one reference work W. In this case, one Mahalanobis distance is calculated for each pixel for all pixels. Here, the Mahalanobis distance for pixel (i, j) is represented as "distance MD(i, j)", and the set of distances MD(i, J) for all pixels is defined as "MD data".

[0063] Similarly, the first feature image generating unit 116 calculates the Mahalanobis distance for all pixels from the array data of pixel values ​​of each pixel for other (second and subsequent) reference workpieces W. This generates multiple MD data for multiple reference workpieces W. In the present disclosure, calculating the Mahalanobis distance for one reference workpiece W based on the data distribution of multiple extracted images CIr obtained from that workpiece W also includes calculation using the average and covariance matrix calculated using data obtained from reference workpieces W other than that workpiece W.

[0064] Returning to Fig. 6 and Fig. 7, next, the control device 100 executes step Sa-5. In step Sa-5, for example, the first feature image generating unit 116 generates a reference feature image DIr from the multiple MD data obtained in step Sa-4. The first feature image generating unit 116 calculates, for each pixel (i, j), a value (pixel value) of that pixel in the reference feature image DIr based on the multiple distances MD(i, j) included in the multiple MD data. The first feature image generating unit 116 may calculate, for each pixel (i, j), the maximum value or average value of the multiple distances MD(i, j) as the pixel value in the reference feature image DIr.

[0065] After calculating the pixel values ​​of the reference feature image DIr for all pixels (i, j), the reference image storage unit 118 stores the reference feature image DIr. With the above, a series of processes in the preparation phase is completed, and a reference feature image DIr, which is standard data used for inspection in the production phase, is generated. In the series of processes exemplified above, one reference feature image DIr is obtained from multiple enhanced images EIr obtained for at least two reference workpieces W. Note that one reference feature image DIr may be obtained from an enhanced image EIr for one reference workpiece W instead of two or more reference workpieces W. Instead of calculating the average and covariance matrix from the array data obtained from all reference workpieces W, the average and covariance matrix may be calculated for each reference workpiece W, and the Mahalanobis distance may be calculated.

[0066] When the two reference workpieces W used to generate the reference feature image DIr are the "reference workpiece Wr1" and the "reference workpiece Wr2", each functional module of the inspection control unit 110 executes the following process. The first input image acquisition unit 112 acquires an enhanced image EIr1 based on an image of the surface Wa of the reference workpiece Wr1, and acquires an enhanced image EIr2 (second reference input image) based on an image of the surface Wa of the reference workpiece Wr2 (second reference substrate). The first intermediate information acquisition unit 114 acquires first intermediate information generated in the intermediate layer when the enhanced image EIr1 is input to the image recognition model M, and acquires second intermediate information (second reference intermediate information) generated in the intermediate layer when the enhanced image EIr2 is input to the image recognition model M. The first feature image generation unit 116 generates the reference feature image DIr based on the first intermediate information and the second intermediate information.

[0067] <Production Phase> Fig. 11 is a flowchart showing a series of processes in the production phase shown in Fig. 6. The series of processes shown in Fig. 11 starts in a state where a workpiece W to be inspected, whose inspection result is unknown, has been transported to the inspection unit U3 after being subjected to a pre-inspection process in the coating and developing process.

[0068] The control device 100 first executes step Sb-1. Step Sb-1 is performed under the same conditions as the processing of step Sa-1 in the preparation phase. In step Sb-1, for example, the second input image acquisition unit 122 of the inspection control unit 110 acquires an image PIs of the surface Wa of the workpiece W to be inspected by imaging the surface Wa of the workpiece W to be inspected using the inspection unit U3.

[0069] Next, the control device 100 executes step Sb-2. Step Sb-2 is performed under the same conditions as the process of step Sa-2 in the preparation phase. In step Sb-2, for example, the second input image acquisition unit 122 generates an enhanced image EIs by performing a process of enhancing contrast on the captured image PIs obtained in step Sb-1.

[0070] Next, the control device 100 executes step Sb-3. Step Sb-3 is performed under the same conditions as the process of step Sa-3 in the preparation phase. In step Sb-3, for example, the second intermediate information acquisition unit 124 acquires an inspection intermediate image generated in the intermediate layer of the image recognition model M when the enhanced image EIs (inspection input image) obtained in step Sb-2 is input to the image recognition model M. The second intermediate information acquisition unit 124 acquires, as inspection intermediate information, an extraction image group CIGs including a plurality of extraction images CIs (a plurality of inspection extracted images) generated based on the enhanced image EIs and the plurality of filters that extract different features in the intermediate layer of the image recognition model M, for example.

[0071] The filters used in generating the extracted images CIs are the same as the filters used in generating the extracted images CIr in the preparation phase. If an arc-shaped scratch is present on the surface Wa of the workpiece W to be inspected, an extracted image CI (feature map) in which the filter reacts to the arc-shaped scratch can be generated in the intermediate layer of the image recognition model M, as in the "extracted image CIj" shown in FIG.

[0072] Next, the control device 100 executes step Sb-4. Step Sb-4 is performed in a manner similar to step Sa-4 in the preparation phase. In step Sb-4, for example, the second feature image generating unit 126 performs a calculation to generate an inspection feature image DIs based on the multiple extraction images CIs acquired in step Sb-3. The second feature image generating unit 126 calculates a Mahalanobis distance (inspection Mahalanobis distance) for array data of pixel values ​​(brightness values) for each pixel (i, j) included in the multiple extraction images CIs, based on the data distribution of multiple extraction images CIr of any one of the extraction image groups CIGr obtained in the preparation phase.

[0073] The second feature image generation unit 126 executes a process of arranging pixel values ​​included in each of the variables x1 to XN corresponding to the N extracted images CIs in a vertical array, similar to step Sa-4 in the preparation phase. The second feature image generation unit 126 calculates the Mahalanobis distance for the horizontally arranged array data of pixel values ​​for each pixel (i, j) included in the N extracted images CIs, using the average and covariance matrix obtained from the multiple extracted images CIr in step Sa-4. In this manner, calculation of the Mahalanobis distance in the present disclosure also includes calculation using the average and covariance matrix used when generating the reference data, rather than the average and covariance matrix obtained from the data to be calculated.

[0074] Next, the control device 100 executes step Sb-5. In step Sb-5, for example, the second feature image generation unit 126 generates an inspection feature image DIs based on the calculation result of the Mahalanobis distance in step Sb-4. For each pixel (i, j), the second feature image generation unit 126 may set the Mahalanobis distance calculated in step Sb-4 to the pixel value of the pixel.

[0075] Next, the control device 100 executes step Sb-6. In step Sb-6, for example, the abnormality determination unit 136 generates a comparison image DiI by comparing the inspection feature image DIs generated in step Sb-5 with the reference feature image DIr stored in the reference image storage unit 118. The abnormality determination unit 136 may calculate the pixel value of each pixel (i, j) in the comparison image DiI by calculating the difference between the pixel value of the inspection feature image DIs and the pixel value of the reference feature image DIr.

[0076] Next, the control device 100 executes step Sb-7. In step Sb-7, for example, the abnormality determination unit 136 determines the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on the result of comparing the inspection feature image DIs with the reference feature image DIr (comparison image DiI generated in step Sb-6). The abnormality determination unit 136 executes a process of extracting pixels having a pixel value equal to or greater than a predetermined value in the inspection feature image DIs. The predetermined value is set to a level at which, when a defect occurs on the surface Wa of the workpiece W, the pixel value at the defective portion can be extracted.

[0077] The abnormality determination unit 136 may determine that there is an abnormality on the surface Wa of the workpiece W to be inspected when an area (or pixel) having a pixel value equal to or greater than a predetermined value is detected in the comparison image DiI. The abnormality determination unit 136 may determine that there is no abnormality on the surface Wa of the workpiece W to be inspected when an area (or pixel) having a pixel value equal to or greater than a predetermined value is not detected in the comparison image DiI.

[0078] Next, the control device 100 executes step Sb-8. In step Sb-8, for example, the judgment result output unit 138 outputs information indicating the judgment result in step Sb-7 to the process control unit 102 or a higher-level controller. When an abnormality signal indicating the presence of an abnormality is output to the process control unit 102 or a higher-level controller, the workpiece W that is determined to have an abnormality (defect) on its surface Wa may be excluded from the processing line following the inspection in the inspection unit U3.

[0079] [Effects of the first embodiment] The substrate inspection method according to the first embodiment described above includes acquiring a reference input image based on an image of the surface Wa of the workpiece W for reference, acquiring reference intermediate information generated in an intermediate layer of the image recognition model M when the reference input image is input to a neural network (image recognition model M) constructed in advance to output a recognition result of the input image, and generating a reference feature image DIr showing the features of the reference input image based on the reference intermediate information. The substrate inspection method further includes acquiring an inspection input image based on an image of the surface Wa of the workpiece W to be inspected, acquiring inspection intermediate information generated in an intermediate layer of the image recognition model M when the inspection input image is input to the image recognition model M, and generating an inspection feature image DIs showing the features of the inspection input image based on the inspection intermediate information. The substrate inspection method further includes determining the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on a result of comparing the reference feature image DIr with the inspection feature image DIs.

[0080] In the intermediate layer of the image recognition model M, a process of extracting a specific shape from an image input to the image recognition model M is executed. Therefore, the reference feature image DIr generated from the information generated in the intermediate layer of the image recognition model M may represent the features of the entire surface Wa of the reference workpiece W. In addition, the inspection feature image DIs generated from the information generated in the intermediate layer of the image recognition model M may represent the features of the entire surface Wa of the workpiece W to be inspected. By comparing the reference feature image DIr and the inspection feature image DIs, the features of the workpiece W itself can be cancelled out, and the features of the abnormal part can be detected. Therefore, the above-mentioned substrate inspection method is useful for accurately detecting an abnormality on the surface Wa of the workpiece W.

[0081] In the above-described substrate inspection method, the reference intermediate information may be a plurality of reference extraction images (a plurality of extraction images CIr) generated in an intermediate layer of the image recognition model M based on a reference input image and a plurality of filters that extract different features from each other. Generating the reference feature image DIr may include generating the reference feature image DIr based on the plurality of extraction images CIr. The inspection intermediate information may be a plurality of inspection extraction images (a plurality of extraction images CIs) generated in an intermediate layer of the image recognition model M based on the inspection input image and the plurality of filters. Generating the inspection feature image DIs may include generating the inspection feature image DIs based on the plurality of extraction images CIs. In this case, various specific shapes are extracted by the plurality of filters in the intermediate layer of the image recognition model M. By extracting various specific shapes, when an abnormal portion is included on the surface Wa of the workpiece W, the abnormal portion can be extracted in response to the filter. Therefore, it is useful for highly accurate abnormality detection on the surface Wa of the workpiece W.

[0082] In the above-described substrate inspection method, generating the reference feature image DIr may include calculating a pixel value for each pixel of the reference feature image DIr based on the result of calculating a reference Mahalanobis distance for array data of pixel values ​​for each pixel (i, j) included in the plurality of extraction images CIr, based on the data distribution of the plurality of extraction images CIr. Generating the inspection feature image DIs may include calculating a pixel value for each pixel of the inspection feature image DIs based on the result of calculating an inspection Mahalanobis distance for array data of pixel values ​​for each pixel (i, j) included in the plurality of extraction images CIs, based on the average and covariance matrix used in calculating the reference Mahalanobis distance. The Mahalanobis distance can represent the degree of deviation (degree of abnormality) from the data distribution. Therefore, if an abnormal portion exists on the surface Wa, the pixel value of a specific pixel will fluctuate in response to the filter in the image recognition model M. This can cause the Mahalanobis distance of a specific pixel to have a large value. In the above configuration, since the feature images are compared with each other, it is possible to detect an area where the Mahalanobis distance is large due to an abnormal part after canceling out the areas where the Mahalanobis distance is large due to the features of the workpiece W itself. Therefore, it is further useful for highly accurate detection of anomalies on the surface Wa of the workpiece W.

[0083] In the substrate processing method described above, the reference input image may be an enhanced image EIr generated by performing a process to enhance contrast on a captured image PIr obtained by capturing the surface Wa of the workpiece W for reference. The inspection input image may be an enhanced image EIs generated by performing a process to enhance contrast on a captured image PIs obtained by capturing the surface Wa of the workpiece W to be inspected. In this case, the part corresponding to the abnormal part is enhanced, and an inspection feature image DIs in which the feature is reflected can be obtained. Even if the part other than the abnormal part is emphasized as noise, the noise can be reduced by comparing the feature images. Therefore, it is further useful for highly accurate detection of abnormalities on the surface Wa of the workpiece W.

[0084] The substrate processing method described above may include acquiring a second reference input image based on an image of the surface Wa of another workpiece W for reference, and acquiring second reference intermediate information generated in an intermediate layer of the image recognition model M when the second reference input image is input to the image recognition model M. Generating the reference feature image DIr may include generating a reference feature image DIr based on the reference intermediate information and the second reference intermediate information. In this case, one reference feature image DIr is generated from the captured images of the surfaces Wa of multiple reference workpieces W. Therefore, the reference feature image DIr can be generated after reducing the influence of the features of one individual workpiece W for reference. Therefore, it is further useful for highly accurate abnormality detection on the surface Wa of the workpiece W.

[0085] [Second embodiment] Next, with reference to Figs. 12 to 16, a substrate processing system 1 according to the second embodiment will be described. In the substrate processing system 1 according to the second embodiment, a series of processes executed by the inspection control unit 110 differs from the series of processes executed by the inspection control unit 110 in the first embodiment. The coating and developing apparatus 2 may execute a coating and developing process (substrate processing) on ​​a predetermined number of workpieces W in lot units. The coating and developing apparatus 2 executes a coating and developing process on a predetermined number of workpieces W in sequence in processing on a lot unit basis (lot processing). The predetermined number indicating a unit of a lot may be determined according to the number of workpieces W that can be accommodated in a carrier C. The coating and developing apparatus 2 repeatedly executes a coating and developing process on a lot unit basis.

[0086] Unlike the inspection method in the first embodiment, the inspection control unit 110 generates a reference feature image, which is reference data used for inspection, while the production phase is being performed. The inspection control unit 110 generates a reference feature image using a workpiece W (first substrate) that is processed first in the coating and developing process on a lot-by-lot basis. In this case, the workpiece W that is processed first becomes a reference workpiece W, although the presence or absence of an abnormality on the surface Wa is unknown. The inspection control unit 110 performs inspection using the reference feature image, with the workpiece W (second substrate) that is processed second or later in the coating and developing process on a lot-by-lot basis being the workpiece W to be inspected.

[0087] Unlike the inspection method in the first embodiment, the inspection control unit 110 executes two different inspection procedures and judges the presence or absence of an abnormality on the surface Wa of the workpiece W from the results obtained from those inspection procedures. FIG. 12 shows a series of processes executed in one inspection procedure, and FIG. 13 shows a series of processes executed in another inspection procedure. In both inspection procedures, a reference feature image is generated from the first workpiece W. FIG. 14 is a flowchart showing an example of a series of processes executed by the inspection control unit 110 on a lot-by-lot basis.

[0088] In this series of processes, the control device 100 executes step 41 in a state where the workpiece W to be processed has been subjected to pre-inspection processing in the coating and developing process and has been transported to the inspection unit U3. In step S41, for example, the inspection control unit 110 determines whether the workpiece W to be processed transported to the inspection unit U3 is the first workpiece W in the processing in lot units. The inspection control unit 110 may determine whether the workpiece W is the first workpiece W in the lot unit by counting the number of workpieces W inspected in the inspection unit U3 from the start of the production phase.

[0089] In step S41, if the workpiece W to be processed that has been transported to the inspection unit U3 is the first workpiece W (step S41: YES), the process executed by the control device 100 proceeds to step S50. In step S50, for example, the inspection control unit 110 executes an inspection preparation process for inspecting the second or subsequent workpieces W to be processed. FIG. 15 is a flowchart showing an example of the inspection preparation process of step S50. The inspection preparation process of step S50 includes a series of processes for the first workpiece shown in FIG. 12 and a series of processes for the first workpiece shown in FIG. 13.

[0090] In the inspection preparation process of step S50, the control device 100 first executes step Sc-1. Step Sc-1 is executed in the same manner as step Sa-1 in the substrate inspection method according to the first embodiment. In step Sc-1, for example, the first input image acquisition unit 112 of the inspection control unit 110 acquires an imaged image PIr of the surface Wa of the first workpiece W by imaging the surface Wa of the first workpiece W using the inspection unit U3.

[0091] Next, the control device 100 executes step Sc-3. Step Sc-3 is executed in the same manner as step Sa-3 in the substrate inspection method according to the first embodiment. In step Sc-3, for example, when the first intermediate information acquisition unit 114 inputs the captured image PIr (reference input image) obtained in step Sc-1 to the image recognition model M, it acquires reference intermediate information generated in the intermediate layer of the image recognition model M. For example, in the intermediate layer of the image recognition model M, the first intermediate information acquisition unit 114 acquires, as reference intermediate information, an extraction image group CIGr1 including a plurality of extraction images CIr1 (a plurality of reference extraction images) generated based on the captured image PIr and a plurality of filters that extract different features from each other. The number of the plurality of extraction images CIr1 included in the extraction image group CIGr1 may be 30 to 60.

[0092] Next, the control device 100 executes step Sc-4. Step Sc-4 is executed in the same manner as step Sa-4 in the substrate inspection method according to the first embodiment. In step Sc-4, for example, the first feature image generation unit 116 calculates a Mahalanobis distance (reference Mahalanobis distance) for array data of pixel values ​​(brightness values) for each pixel (i, j) included in the multiple extraction images CIr1, based on the data distribution of the multiple extraction images CIr1.

[0093] Next, the control device 100 executes step Sc-5. Step Sc-5 is executed similarly to step Sa-5 or step Sb-5 in the substrate inspection method according to the first embodiment. In step Sc-5, for example, the first feature image generation unit 116 generates a reference feature image DIr1 based on the calculation result of the Mahalanobis distance in step Sc-4. For each pixel (i, j), the first feature image generation unit 116 may set the Mahalanobis distance calculated in step Sc-4 to the pixel value of the pixel in the reference feature image DIr1. The reference image storage unit 118 stores the reference feature image DIr1.

[0094] In parallel with or after the series of processes including steps Sc-3, Sc-4, and Sc-5, the control device 100 executes step Se-2. Step Se-2 is executed in the same manner as step Sa-2 in the substrate inspection method according to the first embodiment. In step Se-2, for example, the first input image acquisition unit 112 generates an enhanced image EIr2 by performing a process of enhancing contrast on the captured image PIr obtained in step Sc-1.

[0095] Next, the control device 100 executes step Se-3. Step Se-3 is executed in the same manner as step Sc-3. In step Se-3, for example, when the first intermediate information acquisition unit 114 inputs the enhanced image EIr2 (reference input image) obtained in step Se-2 to the image recognition model M, it acquires reference intermediate information generated in the intermediate layer of the image recognition model M. For example, in the intermediate layer of the image recognition model M, the first intermediate information acquisition unit 114 acquires, as reference intermediate information, an extracted image group CIGr2 including a plurality of extracted images CIr2 (a plurality of second reference extracted images) generated based on the enhanced image EIr2 and a plurality of filters (a plurality of second filters) that extract different features. The number of the extracted images CIr2 included in the extracted image group CIGr2 may be different from the number of the extracted images CIr1 included in the extracted image group CIGr1 obtained in step Sc-3, and may be 180 to 220. That is, the number of the filters for generating the extracted image group may be different between step Sc-3 and step Se-3.

[0096] Next, the control device 100 executes step Se-4. Step Se-4 is executed in the same manner as step Sc-4. In step Se-4, for example, the first feature image generation unit 116 calculates a Mahalanobis distance (reference Mahalanobis distance) for array data of pixel values ​​(luminance values) for each pixel (i, j) included in the multiple extraction images CIr1, based on the data distribution of the multiple extraction images CIr1.

[0097] Next, the control device 100 executes step Se-5. Step Se-5 is executed in the same manner as step Sc-5 described above. In step Se-5, for example, the first feature image generation unit 116 generates a reference feature image DIr2 (second reference feature image) based on the calculation result of the Mahalanobis distance in step Se-4. For each pixel (i, j), the first feature image generation unit 116 may set the Mahalanobis distance calculated in step Se-4 to the pixel value of that pixel in the reference feature image DIr2. The reference image storage unit 118 stores the reference feature image DIr2.

[0098] In this manner, the inspection preparation process in step S50 is completed, and the reference feature images DIr1 and DIr2, which are the reference data used in the inspection of the second and subsequent workpieces W, are generated.

[0099] After executing step S50, the control device 100 executes step S60 as shown in Fig. 14. In step S60, for example, the abnormality determination unit 136 determines the presence or absence of an abnormality on the surface Wa of the first workpiece W based on at least one of the reference feature image DIr1 and the reference feature image DIr2 (for example, the reference feature image DIr2). The abnormality determination unit 136 may execute a process of extracting pixels having a pixel value equal to or greater than a predetermined value in the reference feature image DIr2, and may determine that there is an abnormality (defect) on the surface Wa when an area (or pixel) having a pixel value equal to or greater than the predetermined value is detected. In step S60, although the sensitivity is low, there are cases where an abnormality on the surface Wa of the first workpiece W can be detected.

[0100] On the other hand, in step S41, if the workpiece W to be processed that is transported to the inspection unit U3 is any of the second and subsequent workpieces W (step S41: NO), the process executed by the control device 100 proceeds to step S70. In step S70, for example, the inspection control unit 110 executes an inspection of the workpiece W to be processed that is the second and subsequent workpieces W. FIG. 16 is a flowchart showing an example of the inspection process of step S70. The inspection process of step S70 includes a series of processes for the second and subsequent workpieces shown in FIG. 12 and a series of processes for the second and subsequent workpieces shown in FIG. 13.

[0101] In the inspection process of step S70, the control device 100 first executes step Sd-1. Step Sd-1 is executed under the same conditions as step Sc-1. In step Sd-1, for example, the second input image acquisition unit 122 of the inspection control unit 110 acquires an image PIs of the workpiece W to be inspected by imaging the surface Wa of the workpiece W to be inspected that is to be processed second or later by the inspection unit U3.

[0102] Next, the control device 100 executes step Sd-3. Step Sd-3 is executed under the same conditions as step Sc-3. In step Sd-3, for example, when the second intermediate information acquisition unit 124 inputs the captured image PIs (reference input image) obtained in step Sd-1 to the image recognition model M, it acquires inspection intermediate information generated in the intermediate layer of the image recognition model M. For example, in the intermediate layer of the image recognition model M, the second intermediate information acquisition unit 124 acquires, as inspection intermediate information, an extraction image group CIGs1 including a plurality of extraction images CIs1 (a plurality of inspection extraction images) generated based on the captured image PIs and a plurality of filters that extract different features. The plurality of filters used in step Sd-3 are the same as the plurality of filters used in step Sc-3.

[0103] Next, the control device 100 executes step Sd-4. Step Sd-4 is executed in the same manner as step Sb-4 in the substrate inspection method according to the first embodiment. In step Sd-4, for example, the second feature image generation unit 126 calculates a Mahalanobis distance (inspection Mahalanobis distance) for array data of pixel values ​​(brightness values) for each pixel (i, j) included in the multiple extraction images CIs1 obtained in step Sd-3, based on the data distribution of the multiple extraction images CIr1 obtained in step Sc-3.

[0104] Next, the control device 100 executes step Sd-5. Step Sd-5 is executed in the same manner as step Sc-5 described above. In step Sd-5, for example, the second feature image generation unit 126 generates an inspection feature image DIs1 based on the calculation result of the Mahalanobis distance in step Sd-4. For each pixel (i, j), the second feature image generation unit 126 may set the Mahalanobis distance calculated in step Sd-4 to the pixel value of that pixel in the inspection feature image DIs1.

[0105] Next, the control device 100 executes step Sd-6. Step Sd-6 is executed in the same manner as step Sb-6 in the substrate inspection method according to the first embodiment. In step Sd-6, for example, the abnormality determination unit 136 generates a comparison image DiI1 by comparing the inspection feature image DIs1 generated in step Sd-5 with the reference feature image DIr1 stored in the reference image storage unit 118. The abnormality determination unit 136 may calculate the pixel value of each pixel (i, j) in the comparison image DiI1 by calculating the difference between the pixel value of the inspection feature image DIs1 and the pixel value of the reference feature image DIr1.

[0106] In parallel with or after the series of processes including steps Sd-3 to Sd-6, the control device 100 executes step Sf-2. Step Sf-2 is executed under the same conditions as step Se-2. In step Sf-2, for example, the second input image acquisition unit 122 generates an enhanced image EIs2 by performing a process of enhancing contrast on the captured image PIs obtained in step Sd-1.

[0107] Next, the control device 100 executes step Sf-3. Step Sf-3 is executed under the same conditions as step Se-3. In step Sf-3, for example, when the second intermediate information acquisition unit 124 inputs the enhanced image EIs2 (inspection input image) obtained in step Sf-2 to the image recognition model M, it acquires inspection intermediate information generated in the intermediate layer of the image recognition model M. For example, in the intermediate layer of the image recognition model M, the second intermediate information acquisition unit 124 acquires, as the inspection intermediate information, an extraction image group CIGs2 including a plurality of extraction images CIs2 (a plurality of second inspection extraction images) generated based on the enhanced image EIs2 and a plurality of filters (a plurality of second filters) that extract different features. The plurality of filters used in step Se-3 and the plurality of filters used in step Sf-3 are the same.

[0108] Next, the control device 100 executes step Sf-4. Step Sf-4 is executed in the same manner as step Se-4. In step Sf-4, for example, the second feature image generating unit 126 calculates the Mahalanobis distance (test Mahalanobis distance) for the array data of pixel values ​​(luminance values) for each pixel (i, j) included in the multiple extracted images CIs2 obtained in step Sf-3, based on the data distribution of the multiple extracted images CIs2. In step Sd-4, the mean and covariance matrix used in calculating the Mahalanobis distance to generate the reference data are used, whereas in step Sf-4, the mean and covariance matrix obtained from itself (the multiple extracted images CIs2) are used.

[0109] Next, the control device 100 executes step Sf-5. Step Sf-5 is executed in the same manner as step Se-5 above. In step Sf-5, for example, the second feature image generation unit 126 generates an inspection feature image DIs2 (second inspection feature image) based on the calculation result of the Mahalanobis distance in step Sf-4. For each pixel (i, j), the second feature image generation unit 126 may set the Mahalanobis distance calculated in step Sf-4 to the pixel value of that pixel in the inspection feature image DIs2.

[0110] Next, the control device 100 executes step Sf-6. Step Sf-6 is executed in the same manner as step Sd-6. In step Sf-6, for example, the abnormality determination unit 136 generates a comparison image DiI2 by comparing the inspection feature image DIs2 generated in step Sf-5 with the reference feature image DIr2 stored in the reference image storage unit 118. The abnormality determination unit 136 may calculate the pixel value of each pixel (i, j) in the comparison image DiI2 by calculating the difference between the pixel value of the inspection feature image DIs2 and the pixel value of the reference feature image DIr2.

[0111] Next, the control device 100 executes step S47. In step S47, for example, the abnormality determination unit 136 determines the presence or absence of an abnormality on the surface Wa of the workpiece W to be processed based on a result of comparing the reference feature image DIr1 with the inspection feature image DIs1 and a result of comparing the reference feature image DIr2 with the inspection feature image DIs2. The abnormality determination unit 136 determines the presence or absence of an abnormality on the surface Wa of the workpiece W to be processed based on the comparison image DiI1 obtained in step Sd-6 and the comparison image DiI2 obtained in step Sf-6.

[0112] In one example, the abnormality determination unit 136 executes a process of extracting pixels having a pixel value equal to or greater than a predetermined value in each of the comparison images DiI1 and DiI2. The abnormality determination unit 136 may determine that there is an abnormality on the surface Wa of the workpiece W to be inspected when an area (or pixel) having a pixel value equal to or greater than a predetermined value is detected in at least one of the comparison images DiI1 and DiI2. The abnormality determination unit 136 may determine that there is no abnormality on the surface Wa of the workpiece W to be inspected when an area (or pixel) having a pixel value equal to or greater than a predetermined value is not detected in both the comparison images DiI1 and DiI2.

[0113] Returning to FIG. 14, after execution of step S60 or step S70, the control device 100 executes step S48. In step S48, for example, the judgment result output unit 138 outputs information indicating the judgment results in steps S47 and S60 to the process control unit 102 or a higher-level controller. The workpiece W that is judged to have an abnormality (defect) on the surface Wa may be excluded from the processing line after inspection in the inspection unit U3. If an abnormality is detected in the first workpiece W in step S60, the inspection control unit 110 may execute step S50 for the second workpiece W to generate reference data from a captured image of the second workpiece W, etc.

[0114] Next, the control device 100 executes step S49. In step S49, for example, the control device 100 determines whether or not inspection of a predetermined number of workpieces W that defines a lot unit has been completed. If it is determined in step S49 that inspection of the predetermined number of workpieces W has not been completed (step S49: NO), the process executed by the control device 100 returns to step S41. If it is determined in step S49 that inspection of the predetermined number of workpieces W has been completed (step S49: YES), the board inspection for one lot ends. The control device 100 (inspection control unit 110) executes a similar board inspection process for the next lot.

[0115] [Effects of the second embodiment] The substrate inspection method performed in the substrate processing system 1 according to the second embodiment also has the same effects as those of the first embodiment, and is therefore useful for detecting abnormalities on the front surface Wa of the workpiece W with high accuracy.

[0116] In the substrate inspection method according to the second embodiment described above, generating the reference feature image DIr2 may include calculating a pixel value for each pixel of the reference feature image DIr2 based on the result of calculating the reference Mahalanobis distance for the array data of pixel values ​​for each pixel included in the plurality of reference extraction images (the plurality of extraction images CIr2) based on the data distribution of the plurality of extraction images CIr2. Generating the inspection feature image DIs2 may include calculating a pixel value for each pixel of the inspection feature image DIs2 based on the result of calculating the inspection Mahalanobis distance for the array data of pixel values ​​for each pixel included in the plurality of inspection extraction images (the plurality of extraction images CIs2) based on the data distribution of the plurality of extraction images CIs2. If the inspection Mahalanobis distance is calculated using the average and covariance matrix used when calculating the reference Mahalanobis distance when creating the reference data, even a change between the reference work W and the work W to be inspected that is not to be determined as an abnormality may be detected. In the above configuration, the inspection Mahalanobis distance is calculated from its own data distribution, and the difference from the reference work W is not reflected in the inspection Mahalanobis distance. Therefore, it is useful to adjust the detection sensitivity according to the type of abnormality to be detected.

[0117] In the substrate inspection method according to the second embodiment described above, the reference workpiece W may be the first workpiece W to be subjected to coating and developing treatment in a lot process in which a predetermined number of workpieces W to be processed are sequentially subjected to a predetermined coating and developing treatment. The workpiece W to be inspected may be any of the second and subsequent workpieces W to be subjected to coating and developing treatment in any order from the second workpiece W in the lot process. In this case, reference data for inspection is generated on a lot-by-lot basis. Therefore, variations in the workpiece W itself or the coating and developing treatment between multiple lots are unlikely to affect abnormality judgment. Therefore, it is useful for highly accurate abnormality detection on the surface Wa of the workpiece W.

[0118] The board inspection method according to the second embodiment described above includes, in addition to generating a reference feature image DIr1 and an inspection feature image DIs1, acquiring, in an intermediate layer of the image recognition model M, a plurality of second reference extraction images (a plurality of extraction images CIr2) generated based on the reference input image and a plurality of second filters that extract features different from each other when a reference input image is input to the image recognition model M, and generating the second reference feature image (reference feature image DIr2) based on the plurality of extraction images CIr2. The board inspection method may further include, in an intermediate layer of the image recognition model M, acquiring, in an intermediate layer of the image recognition model M, a plurality of second inspection extraction images (a plurality of extraction images CIs2) generated based on the inspection input image and the plurality of second filters when an inspection input image is input to the image recognition model M, and generating the second inspection feature image (inspection feature image DIs2) based on the plurality of extraction images CIs2 when an inspection input image is input to the image recognition model M. Determining the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected may include determining the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on the result of comparing the reference feature image DIr1 with the inspection feature image DIs1 and the result of comparing the reference feature image DIr2 with the inspection feature image DIs2. In this case, since the feature images are compared in two different inspection procedures, it is possible to reduce the possibility of not being able to detect an abnormality while facilitating adjustment of the detection sensitivity according to the type of abnormality to be detected. Therefore, it is useful for achieving both adjustment of the detection sensitivity for an abnormality and detection accuracy.

[0119] The substrate inspection method according to the second embodiment described above may further include judging the presence or absence of an abnormality on the surface Wa of the first workpiece W based on the reference feature image. In this case, if an obvious abnormality is included in the surface Wa of the first workpiece W for generating the reference data for inspection, that workpiece W can be rejected, and the reference data can be created using another workpiece W. Therefore, it is useful for highly accurate detection of an abnormality on the surface Wa of the workpiece W.

[0120] [Variations] The present disclosure is not limited to the first and second embodiments described above. Some of the matters described in the first embodiment may be applied to the second embodiment, and some of the matters described in the second embodiment may be applied to the first embodiment. The above-mentioned series of processes can be modified as appropriate. In the above series of processes, the control device 100 (inspection control unit 110) may execute one step and the next step in parallel, or may execute each step in an order different from the above-mentioned example. The control device 100 (inspection control unit 110) may omit any step, or may execute a process different from the above-mentioned example in any step.

[0121] The inspection control unit 110 according to the first embodiment may omit steps Sa-2 and Sb-2. In this case, the inspection control unit 110 inputs the captured image PIr to the image recognition model M in step Sa-3, and inputs the captured image PIs to the image recognition model M in step Sb-3. The inspection control unit 110 according to the first embodiment may generate a reference feature image DIr, which is standard data, from one reference workpiece W.

[0122] The inspection control unit 110 according to the second embodiment executes two different inspection procedures, but may execute only one of the inspection procedures. The inspection control unit 110 may determine the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on a comparison image DiI1 obtained by executing the inspection procedure shown in Fig. 12 without executing the inspection procedure shown in Fig. 13. The inspection control unit 110 may determine the presence or absence of an abnormality on the surface Wa of the workpiece W to be inspected based on a comparison image DiI2 obtained by executing the inspection procedure shown in Fig. 13 without executing the inspection procedure shown in Fig. 12.

[0123] In the inspection procedure according to the second embodiment, a series of processes performed on the first workpiece W may be executed in a preparation phase before the start of the production phase. Then, in the production phase, a series of processes performed on the second and subsequent workpieces W may be executed on the workpieces W to be processed regardless of the lot unit.

[0124] The inspection control unit 110 may not hold the image recognition model M, and an external device of the control device 100 may store the image recognition model M. In this case, the inspection control unit 110 may transmit the captured image or the highlighted image to the external device, and then acquire the extracted image group CIG from the external device.

[0125] A computer constituting the inspection control unit 110 may be provided outside the coating and developing apparatus 2. In this case, the control device 100 and the inspection control unit 110 may be connected to be able to communicate with each other via a wire, wirelessly, or a network. The control device 100 may acquire a captured image from the inspection unit U3 and transmit the captured image to the inspection control unit 110. The inspection control unit 110 may transmit information indicating a determination result as to whether or not there is an abnormality to the control device 100. At least a part of the matters described in the other examples may be applied to one example among the various examples described above. [Explanation of symbols]

[0126] 1...substrate processing system, 2...coating and developing apparatus, U3...inspection unit, W...work, 110...inspection control unit, 112...first input image acquisition unit, 114...first intermediate information acquisition unit, 116...first feature image generation unit, 122...second input image acquisition unit, 124...second intermediate information acquisition unit, 126...second feature image generation unit, 136...abnormality judgment unit, PIr, PIs...captured image, EIr, EIs...enhanced image, CIr, CIs...extracted image, DIr...reference feature image, DIs...inspection feature image, DiI...comparison image.

Claims

1. Obtaining a reference input image based on an image captured of a surface of a reference substrate; When the reference input image is input to a neural network that is constructed in advance to output a recognition result of an input image, a plurality of reference extracted images are obtained in an intermediate layer of the neural network, the reference input image being generated based on a plurality of filters that extract different features from each other; generating a reference feature image indicating features of the reference input image based on the plurality of reference extracted images; acquiring an inspection input image based on an image captured of a surface of a substrate to be inspected; acquiring, when the test input image is input to the neural network, a plurality of test extraction images that are generated based on the test input image and the plurality of filters in an intermediate layer of the neural network; generating an inspection feature image indicating features of the inspection input image based on the plurality of inspection extraction images; and determining the presence or absence of an abnormality on the surface of the substrate to be inspected based on a result of comparing the reference feature image with the inspection feature image.

2. generating the reference feature image includes calculating a pixel value for each pixel of the reference feature image based on a result of calculating a reference Mahalanobis distance for array data of pixel values ​​for each pixel included in the plurality of reference extracted images, based on a data distribution of the plurality of reference extracted images; 2. The substrate inspection method of claim 1, wherein generating the inspection feature image includes calculating a pixel value for each pixel of the inspection feature image based on a result of calculating an inspection Mahalanobis distance for array data of pixel values ​​for each pixel included in the multiple inspection extraction images, based on an average and covariance matrix used when calculating the reference Mahalanobis distance.

3. generating the reference feature image includes calculating a pixel value for each pixel of the reference feature image based on a result of calculating a reference Mahalanobis distance for array data of pixel values ​​for each pixel included in the plurality of reference extracted images, based on a data distribution of the plurality of reference extracted images; 2. The substrate inspection method according to claim 1, wherein generating the inspection feature image includes calculating a pixel value for each pixel of the inspection feature image based on a result of calculating an inspection Mahalanobis distance for array data of pixel values ​​for each pixel contained in the plurality of inspection extraction images, based on a data distribution of the plurality of inspection extraction images.

4. the reference input image is an image generated by performing a process for enhancing contrast on an image obtained by capturing an image of a surface of the reference substrate, The substrate inspection method according to any one of claims 1 to 3, wherein the inspection input image is an image generated by performing a process to enhance contrast on an image captured of the surface of the substrate to be inspected.

5. acquiring a second reference input image based on an image captured of a surface of a second substrate for reference; acquiring, in an intermediate layer of the neural network, a plurality of second reference extracted images generated based on the second reference input image and the plurality of filters when the second reference input image is input to the neural network, A substrate inspection method according to any one of claims 1 to 3, wherein generating the reference feature image includes generating the reference feature image based on the plurality of reference extraction images and the plurality of second reference extraction images.

6. the reference substrate is a first substrate that is to be subjected to a predetermined substrate treatment in a lot process in which a predetermined number of substrates to be processed are sequentially subjected to the substrate treatment, 4. The substrate inspection method according to claim 1, wherein the substrate to be inspected is a second substrate which is subjected to the substrate processing in any order from the second substrate onward in the lot processing.

7. the reference substrate is a first substrate that is to be subjected to a predetermined substrate treatment in a lot process in which a predetermined number of substrates to be processed are sequentially subjected to the substrate treatment, the substrate to be inspected is a second substrate to be subjected to the substrate processing in any order from the second substrate onward in the lot processing, The substrate inspection method includes: When the reference input image is input to the neural network, a plurality of second reference extracted images are obtained in an intermediate layer of the neural network based on the reference input image and a plurality of second filters that extract different features from each other; generating a second reference feature image based on the plurality of second reference extracted images; acquiring, when the test input image is input to the neural network, a plurality of second test extraction images generated based on the test input image and the plurality of second filters in an intermediate layer of the neural network; generating a second inspection feature image based on the plurality of second inspection extracted images; A substrate inspection method according to any one of claims 1 to 3, wherein determining the presence or absence of an abnormality on the surface of the substrate to be inspected includes determining the presence or absence of an abnormality on the surface of the second substrate based on a result of comparing the reference feature image with the inspection feature image and a result of comparing the second reference feature image with the second inspection feature image.

8. The substrate inspection method according to claim 6 , further comprising: determining the presence or absence of an abnormality on the surface of the first substrate based on the reference feature image.

9. The substrate inspection method according to claim 7 , further comprising: determining the presence or absence of an abnormality on the surface of the first substrate based on the reference feature image.

10. A substrate inspection program for causing a computer to execute the substrate inspection method according to any one of claims 1 to 3.

11. a first input image acquisition unit that acquires a reference input image based on an image of a surface of a reference substrate; a first intermediate information acquisition unit that acquires, when a reference input image is input to a neural network that is constructed in advance to output a recognition result of an input image, a plurality of reference extracted images that are generated based on the reference input image and a plurality of filters that extract different features in an intermediate layer of the neural network; a first feature image generating unit that generates a reference feature image indicating a feature of the reference input image based on the plurality of reference extracted images; a second input image acquisition unit that acquires an inspection input image based on an image of a surface of a substrate to be inspected; a second intermediate information acquiring unit that acquires, when the test input image is input to the neural network, a plurality of test extraction images generated based on the test input image and the plurality of filters in an intermediate layer of the neural network; a second feature image generating unit that generates an inspection feature image indicating a feature of the inspection input image based on the plurality of inspection extraction images; A substrate inspection apparatus comprising: an abnormality determination unit that determines the presence or absence of an abnormality on the surface of the substrate to be inspected based on a result of comparing the reference feature image with the inspection feature image.

12. In generating the reference feature image, the first feature image generating unit calculates a pixel value for each pixel of the reference feature image based on a result of calculating a reference Mahalanobis distance for array data of pixel values ​​for each pixel included in the plurality of reference extraction images, based on a data distribution of the plurality of reference extraction images; The substrate inspection apparatus of claim 11, wherein the second feature image generation unit, in generating the inspection feature image, calculates a pixel value for each pixel of the inspection feature image based on a result of calculating an inspection Mahalanobis distance based on an average and covariance matrix used when calculating the reference Mahalanobis distance for array data of pixel values ​​for each pixel contained in the multiple inspection extraction images.

13. In generating the reference feature image, the first feature image generating unit calculates a pixel value for each pixel of the reference feature image based on a result of calculating a reference Mahalanobis distance for array data of pixel values ​​for each pixel included in the plurality of reference extraction images, based on a data distribution of the plurality of reference extraction images; The substrate inspection apparatus of claim 11, wherein the second feature image generation unit, in generating the inspection feature image, calculates pixel values ​​for each pixel of the inspection feature image based on a result of calculating an inspection Mahalanobis distance for array data of pixel values ​​for each pixel contained in the plurality of inspection extraction images, based on a data distribution of the plurality of inspection extraction images.

14. The first input image acquisition unit generates the reference input image by performing a process for enhancing contrast on an image of a surface of the reference substrate, The substrate inspection apparatus according to any one of claims 11 to 13, wherein the second input image acquisition unit generates the inspection input image by performing a process to enhance contrast on an image captured of the surface of the substrate to be inspected.

15. The first input image acquisition unit further acquires a second reference input image based on an image of a surface of a second substrate for reference; a first intermediate information acquisition unit, when the second reference input image is input to the neural network, further acquires a plurality of second reference extracted images generated based on the second reference input image and the plurality of filters in an intermediate layer of the neural network; The substrate inspection apparatus according to any one of claims 11 to 13, wherein the first feature image generation unit generates the reference feature image based on the plurality of reference extraction images and the plurality of second reference extraction images.

16. The reference substrate is a first substrate that is to be subjected to a predetermined substrate treatment in a lot process in which a predetermined number of substrates to be processed are sequentially subjected to the substrate treatment, 14. The substrate inspection apparatus according to claim 11, wherein the substrate to be inspected is a second substrate which is subjected to the substrate processing in any order from the second substrate onward in the lot processing.

17. The reference substrate is a first substrate that is to be subjected to a predetermined substrate treatment in a lot process in which a predetermined number of substrates to be processed are sequentially subjected to the substrate treatment, the substrate to be inspected is a second substrate to be subjected to the substrate processing in any order from the second substrate onward in the lot processing, the first intermediate information acquisition unit, when inputting the reference input image to the neural network, further acquires, in an intermediate layer of the neural network, a plurality of second reference extracted images generated based on the reference input image and a plurality of second filters that extract features different from each other; the first feature image generating unit further generates a second reference feature image based on the plurality of second reference extracted images; the second intermediate information acquisition unit, when inputting the test input image to the neural network, further acquires a plurality of second test extraction images generated based on the test input image and the plurality of second filters in an intermediate layer of the neural network; the second feature image generating unit further generates a second inspection feature image based on the plurality of second inspection extraction images; A substrate inspection apparatus as described in any one of claims 11 to 13, wherein the abnormality judgment unit, in determining the presence or absence of an abnormality on the surface of the substrate to be inspected, judges the presence or absence of an abnormality on the surface of the second substrate based on a result of comparing the reference feature image with the inspection feature image and a result of comparing the second reference feature image with the second inspection feature image.

18. A substrate inspection apparatus as described in Claim 16, wherein the abnormality judgment unit further judges the presence or absence of an abnormality on the surface of the first substrate based on the reference feature image.

19. A substrate inspection apparatus as described in Claim 17, wherein the abnormality judgment unit further judges the presence or absence of an abnormality on the surface of the first substrate based on the reference feature image.

Images