Inspection apparatus

Abstract

To provide an inspection device capable of quickly and accurately performing inspection even when performing the inspection using deep learning.SOLUTION: When pixel information serving as a reference of a non-defective product is stored, images are acquired from a plurality of non-defective product images, and statistical processing of variation in luminance for each pixel is performed on the images. Then, a specific region 71 having a large standard deviation with respect to the luminance of the pixel is extracted. Next, when the inspection is performed, the inspection is executed by the comparison inspection for the area other than the specific area 71, and the inspection is performed by the deep learning for the specific area 71 for the inspection object determined as the "non-defective product".SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to an inspection device that uses deep learning to inspect the quality of the formation state of an object to be inspected. [Background technology]

[0002] In recent years, methods have been proposed for inspecting the quality of inspection objects using deep learning.

[0003] For example, Patent Document 1 listed below proposes a method for acquiring an image from an object to be inspected and applying deep learning to the acquired image, thereby enabling high-speed and accurate inspection. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-105758 Summary of the Invention [Problem to be solved by the invention]

[0005] However, when performing inspection using such deep learning, the following problems arise.

[0006] In other words, deep learning requires large-scale calculations, and therefore requires sufficient time and resources when training images with a large number of pixels. For this reason, in practice, image information is thinned out to create smaller images for training. However, when inspections are performed using such thinning, image information is lost, resulting in poor detection accuracy. Another method is to divide the inspection area into smaller units for training, but this requires verifying the reliability of the divided areas and also poses problems such as difficulty in detecting global anomalies.

[0007] Therefore, the present invention has been made with a focus on the above-mentioned problems, and aims to provide an inspection device that enables quick and accurate inspection even when inspection is performed using deep learning. [Means for solving the problem]

[0008] That is, in order to solve the above-mentioned problems, the present invention comprises an image acquisition means for acquiring images from a plurality of non-defective articles that are deemed to be non-defective, a storage means for storing pixel information that serves as a standard for non-defective articles based on the images acquired by the image acquisition means, a specific area extraction means for performing statistical processing of brightness variations for each pixel of the images stored in the storage means and extracting specific areas with a large pixel standard deviation from the pixels obtained by the statistical processing, a first inspection means for inspecting the quality of the object to be inspected by comparing the image acquired by the image acquisition means with the standard pixel information stored in the storage means, and a second inspection means for inspecting the specific areas extracted by the specific area extraction means using deep learning based on the images acquired by the image acquisition means.

[0009] In such an invention, a comparison inspection is performed using the first inspection means on areas other than the specific area extracted by the specific area extraction means. [Effects of the Invention]

[0010] According to the present invention, deep learning is used to inspect only specific areas with large variations obtained from multiple images of good products, thereby shortening the learning time and enabling highly accurate inspection. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a schematic diagram of an inspection device according to an embodiment of the present invention; [Figure 2] Functional block diagram of the same configuration [Figure 3] FIG. 10 is a diagram showing a memory map stored in a storage means of an object to be inspected in the same embodiment. [Figure 4] Figure showing pixel-by-pixel statistical information and specific regions in the same format [Figure 5] Flowchart showing the inspection process in the same embodiment DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0013] As shown in FIGS. 1 and 2 , the inspection device 1 of this embodiment acquires an image of the surface of an inspection object 7 and compares it with image information acquired from a non-defective product to inspect the quality of the inspection object 7. The inspection device 1 includes an image acquisition means 2 for acquiring an image from the inspection object 7, a storage means 3 for storing information based on images of non-defective products in advance, and an inspection means (5, 6) for inspecting the quality of the inspection object 7 using the images acquired by the image acquisition means 2 and information based on the images of non-defective products stored in the storage means 3. Characteristically, when creating reference information from images of non-defective products in advance, multiple images of non-defective products are acquired and a specific region 71 having pixels with large luminance variations is set. During inspection, regions other than the specific region 71 are compared to determine whether the pixels have luminance within a certain range relative to the luminance of the reference pixels to determine the region of defective pixels. Meanwhile, for specific regions with large luminance variations, AI inspection using deep learning is performed to determine the quality of the region. The inspection device 1 of this embodiment will be described in detail below.

[0014] First, as shown in Fig. 1, the image acquisition means 2 irradiates the inspection object 7 with light from the light emitting device 21 and acquires an image using the light receiving device 22 that receives the reflected light. When acquiring images with this image acquisition means 2, a plurality of articles that have been judged to be "good" by visual inspection or the like are prepared in advance, and images are acquired from each of the articles. At this time, the respective images are aligned, and the brightness of each aligned pixel is acquired and stored in the storage means 3.

[0015] When storing information about "good" images in the storage means 3, statistical quantities such as the average luminance, standard deviation, maximum luminance, and minimum luminance are extracted for each aligned pixel based on multiple good-quality images acquired by the image acquisition means 2. Then, a rule such as "average ±3 × standard deviation" is applied to the statistical quantities for each pixel coordinate to determine a luminance range, such as an upper limit and a lower limit, for each pixel that is considered a good product. Then, as shown in FIG. 3, the luminance range, such as the upper limit and lower limit, and the standard deviation are stored for each pixel. When storing this pixel luminance information, if a color inspection is performed, 256 levels of luminance information for each RGB are stored, and if variation is to be stored, the standard deviation is stored for each RGB.

[0016] As shown in Fig. 4, the specific region extraction means 4 sets a specific region 71 (region surrounded by a dashed line in Fig. 4) with large variations based on the standard deviation of each pixel in an image acquired from a non-defective product. Specifically, a reference value is set for the standard deviation, and pixels having a standard deviation that deviates from the reference value are extracted. If multiple such pixels exist adjacent to each other, the region including these pixels is set as the specific region 71. If the surface of an object is uneven, a possible example of such a dispersion region would be the boundary between the unevenness.

[0017] The inspection means (5, 6) inspects the images acquired by the image acquisition means 2 using the reference information stored in the memory means 3, and performs the inspection using two types of inspection methods: a first inspection means 5 and a second inspection means 6.

[0018] First, the first inspection means 5 determines whether the object is "good" or "bad" by comparing it with an image of a good product. Specifically, it aligns the image of the object 7 under inspection with the image of a good product, and determines whether the pixels of the object 7 under inspection fall within the range of upper and lower luminance limits of the reference pixel stored in the storage means 3. If the pixels of the object 7 under inspection fall within the range of luminance between the upper and lower luminance limits, they are determined to be "good," and if they fall outside this range, they are determined to be "defective pixels." If "defective pixels" exist adjacent to each other, the adjacent pixels are grouped together into particles, and if the area of ​​the particles exceeds a certain area, the area of ​​the pixels is determined to be a "defective portion."

[0019] On the other hand, the second inspection means 6 performs inspection of a specific region 71 of the inspection object 7 using deep learning.

[0020] Specifically, an image of a specific region 71 of the inspection object 7 is input to the input layer, and the data is weighted in each layer of the multi-layer structure and output through an activation function. The output value is then input to the next layer, where it is again weighted and subjected to the activation function, and input and output are performed sequentially, until a "pass / fail" result for the image is output.

[0021] In this way, the first inspection means 5 using comparative inspection and the second inspection means 6 using deep learning are used to inspect the inspection object 7.

[0022] Next, an inspection method in the inspection device 1 configured as above will be described with reference to the flowcharts of FIGS.

[0023] First, prior to the inspection, a plurality of articles that have been judged to be non-defective by visual inspection are prepared, and an image is acquired from each of the articles by the image acquisition means 2 (step S1).

[0024] Then, the acquired images are aligned with each other using a plurality of reference regions (step S2).

[0025] Then, brightness information is obtained from each pixel of the aligned images (step S3), and statistics such as the average brightness, standard deviation, maximum brightness, and minimum brightness of each pixel are obtained (step S4). This process is performed for all non-defective products, and the brightness and statistical information of each pixel are stored in the storage means 3.

[0026] Next, based on the image thus stored in the storage means 3, pixels whose standard deviation is greater than a certain value are extracted for each pixel (step S5), and if there are multiple adjacent pixels, the pixels are grouped together into particles and set as a specific region 71 (step S6).

[0027] In this way, after pixel information for non-defective products and specific area 71 are set, the quality of inspection object 7 is inspected (FIG. 6).

[0028] In the inspection, an image of the inspection object 7 is acquired from the image acquisition means 2 (step T1), and the acquired image is aligned with the image stored in the storage means 3 (step T2).

[0029] Then, for the image of the aligned inspection object 7, the areas other than the specific area 71 are compared with the pixel information stored in the memory means 3 (step T3), and if each pixel falls within a certain brightness range for the pixels stored in the memory means 3 (step T4: Yes), subsequent inspection using deep learning is performed (step T9).

[0030] On the other hand, if the brightness is outside the predetermined range in step T4 (step T4: No), it is judged to be a "bad pixel" (step T5), and if there are multiple adjacent "bad pixels," they are grouped together and granulated (step T6). If the area exceeds a predetermined reference value (step T7), the product is judged to be "defective" (step T8), and if it is within the predetermined reference value, inspection using deep learning is performed in step T9.

[0031] Next, deep learning in the specific area 71 is inspected (step T9).

[0032] If this deep learning inspection determines that the pixels in the specific area 71 are "good" (step T9: OK), then the object 7 to be inspected is ultimately determined to be a "good product" and this is output (step T10). Conversely, if the area is determined to be "bad," then the object 7 to be inspected is ultimately determined to be a "bad product" (step T8) and the process moves on to visual inspection, etc.

[0033] As described above, according to the embodiment, the apparatus is provided with an image acquisition means 2 that acquires images from a plurality of non-defective articles that are deemed to be non-defective, a storage means 3 that stores pixel information that serves as a standard for non-defective articles based on the images acquired by the image acquisition means 2, a specific area extraction means 4 that performs statistical processing of the brightness variation for each pixel of the images stored in the storage means 3 and extracts specific areas 71 with large pixel deviations from the pixels obtained by the statistical processing, a first inspection means 5 that compares the image of the inspection object 7 acquired by the image acquisition means 2 with the standard pixel information stored in the storage means 3 to inspect the pass / fail of the object, and a second inspection means 6 that inspects the specific areas 71 extracted by the specific area extraction means 4 using deep learning based on the images acquired by the image acquisition means 2. This makes it possible to shorten the learning time for specific areas 71 with large variations and to perform accurate inspections.

[0034] The present invention is not limited to the above-described embodiment, but can be implemented in various forms.

[0035] For example, in the above embodiment, the first inspection means 5 performs a comparative inspection of the brightness of each pixel to determine whether the object 7 is good or bad, but other inspection methods may also be used for inspection.

[0036] Furthermore, in the above embodiment, the first inspection means 5 first performs a comparative inspection of the entire object 7, and then the second inspection means 6 is used to inspect only those objects 7 for which only the specific area 71 is judged to be "defective." However, even for objects 7 that are judged to be "good" by the first inspection means 5, the second inspection means 6 may be used to inspect the specific area 71 using deep learning.

[0037] Furthermore, in the above embodiment, a comparative inspection is performed outside the specific region 71, and an inspection by deep learning is performed within the specific region 71, but a comparative inspection may be performed for all regions including the specific region 71, and a separate inspection by deep learning may be performed for the specific region 71. Then, both the comparative inspection and the inspection by deep learning for the specific region 71 may be output, and verification may be performed based on the respective inspection results. [Explanation of symbols]

[0038] 1. Inspection equipment 2. Image acquisition method 3...Storage means 4... Specific area extraction means 5. First inspection method 6. Second inspection method 7. Inspection object 71...specific area

Claims

1. image acquisition means for acquiring images from a plurality of non-defective articles; a storage means for storing pixel information serving as a standard for a non-defective product based on the image acquired by the image acquisition means; a specific region extraction means for performing statistical processing of luminance variations for each pixel of the image stored in the storage means, and extracting a specific region having a large standard deviation of pixels from among the pixels obtained by the statistical processing; a first inspection means for inspecting the quality of an object by comparing the image of the object acquired by the image acquisition means with reference pixel information stored in the storage means; A second inspection means for inspecting the specific area extracted by the specific area extraction means using deep learning based on the image acquired by the image acquisition means; An inspection device comprising:

2. 2. The inspection apparatus according to claim 1, wherein the first inspection means inspects an area other than the specific area extracted by the specific area extraction means.

3. acquiring images from a plurality of good articles that are good; storing pixel information that serves as a standard for a non-defective product based on the acquired image; performing statistical processing of luminance variations for each pixel of the stored image, and extracting a specific region having a large pixel standard deviation from the pixels obtained by the statistical processing; a step of comparing the acquired image of the object to be inspected with the stored reference pixel information to inspect whether the image is good or bad; Inspecting the extracted specific region by deep learning based on the acquired image of the inspection object; An inspection method comprising:

4. acquiring images from a plurality of good articles that are good; storing pixel information that serves as a standard for a non-defective product based on the acquired image; performing statistical processing of luminance variations for each pixel of the stored image, and extracting a specific region having a large pixel standard deviation from the pixels obtained by the statistical processing; a step of comparing the acquired image of the inspection object with the stored reference pixel information for an area other than the extracted specific area, and inspecting the entire inspection object; A step of inspecting the extracted specific region using deep learning; An inspection method comprising:

Images