JP7118365B2 - Image inspection device - Google Patents

Description

The present invention relates to an image inspection apparatus.

2. Description of the Related Art Conventionally, an image inspection apparatus has been used that takes an image of an object such as a part or a finished product and inspects the object by image processing. Imaging inspection equipment may include a classifier that identifies whether an object is normal or abnormal based on the image.

For example, the following patent document 1 includes M types of convolutional neural network classifiers that are different from each other, the first level convolutional neural network classifier is used for object classification, and the second level convolutional neural network classifier is used to classify objects. A convolutional neural network classifier system is described that is used to classify objects that are difficult to classify for a first level convolutional neural network classifier.

In addition, in Patent Document 2 below, using a large classification neural network and a fine classification neural network, the imaged character is subjected to large classification processing, and if it is specified that the imaged character cannot be specified, Then, a recognition method using a neural network is described in which a process of specifying imaged characters is performed from the next most likely group, and the imaged characters are specified while repeating this process.

JP 2014-49118 A JP-A-7-306916

As in the prior art, a classifier that identifies an object based on an image may be configured by a neural network. However, a discriminator using a neural network may have a relatively large computational load.

Here, the images of the object may include images that are clearly abnormal and images that are clearly normal. Even for images, high-load processing may be uniformly performed, and the time required for determination may be long.

Accordingly, the present invention provides an image inspection apparatus using a neural network that can shorten the expected value of the time required for determination.

An image inspection apparatus according to an aspect of the present disclosure uses an imaging unit that captures an image of an object, and a first neural network and a second neural network each including a plurality of neurons to determine the state of the object based on the image. and a determination unit that determines the first neural network, compared to the second neural network, the calculation load is light, the false negative rate is equal to or lower than the same, and the false positive rate is large. When the result of determining the state of the object using the first neural network satisfies a predetermined condition, the state of the object is determined using the first neural network, and the state of the object is determined using the first neural network. If the result does not satisfy a predetermined condition, a second neural network is used to determine the state of the object.

According to this aspect, by omitting the determination by the second neural network when the determination result by the first neural network whose computational load is lighter than that of the second neural network satisfies the predetermined condition, the expected value of the time required for the determination. can be shortened.

The above aspect may further include a generating unit that generates the second neural network by learning processing using a plurality of learning images and generates the first neural network by reducing the scale of the second neural network.

According to this aspect, it is possible to efficiently generate the first neural network smaller in scale than the second neural network.

In the above aspect, the generation unit fixes one or more neurons on the input layer side among the plurality of neurons included in the hidden layer of the second neural network, and sets the parameters of the one or more neurons on the output layer side to zero. , the first neural network may be generated by performing learning processing using a plurality of learning images.

According to this aspect, the first neural network can be generated by pruning the hidden layer neurons of the second neural network and re-learning, and the first neural network has a lighter computational load than the second neural network. can be generated efficiently.

In the above aspect, the generating unit may generate the first neural network such that a scale ratio of the first neural network to the second neural network is equal to or less than a predetermined value.

According to this aspect, it is possible to make the memory area occupied by the first neural network less than or equal to the predetermined ratio of the memory area occupied by the second neural network.

In the above aspect, the generation unit generates a plurality of candidate neural networks by reducing the scale of the second neural network through a plurality of different processes, and determines the object based on the image by the plurality of candidate neural networks and the second neural network. A computation time to determine the state may be measured and a first neural network may be selected from the plurality of candidate neural networks based on the computation time.

According to this aspect, the first neural network can be selected so that the expected value of the time required for determination by the first neural network and the second neural network is the shortest.

In the above aspect, the determination unit temporarily stores the value of a predetermined neuron among a plurality of neurons included in the hidden layer of the first neural network, and determines the state of the object using the first neural network. is not satisfied, the state of the object may be determined by the second neural network using the temporarily stored predetermined neuron values.

According to this aspect, by sharing the value calculated by the first neural network with the second neural network, the time required for determination by the second neural network can be shortened.

In the above aspect, the generation unit may generate the first neural network such that a ratio of predetermined neurons to a plurality of neurons included in the hidden layer of the first neural network is equal to or greater than a predetermined value.

According to this aspect, by setting the ratio of neurons of the second neural network that share the value calculated by the first neural network to a predetermined value or more, the time required for determination by the second neural network can be shortened by a predetermined time or more. can be done.

According to the present invention, it is possible to provide an image inspection apparatus using a neural network that can shorten the expected value of the time required for determination.

1 is a diagram showing an overview of an image inspection apparatus according to an embodiment of the present invention; FIG. It is a figure which shows the functional block of the information processing apparatus which concerns on this embodiment. It is a figure which shows the physical structure of the information processing apparatus which concerns on this embodiment. It is a figure which shows an example of the image image|photographed by the image inspection apparatus which concerns on this embodiment. It is a figure which shows an example of the 2nd neural network used by the image inspection apparatus which concerns on this embodiment. It is a figure which shows an example of the 1st neural network used by the image inspection apparatus which concerns on this embodiment. It is a figure which shows an example of several candidate neural networks produced|generated by the image inspection apparatus which concerns on this embodiment. 4 is a flowchart of first neural network selection processing executed by the image inspection apparatus according to the present embodiment; 4 is a flowchart of state determination processing executed by the image inspection apparatus according to the present embodiment;

Embodiments of the present invention will be described with reference to the accompanying drawings. It should be noted that, in each figure, the same reference numerals have the same or similar configurations.

FIG. 1 is a diagram showing an overview of an image inspection apparatus 100 according to an embodiment of the invention. The image inspection apparatus 100 includes an information processing device 10 , an illumination section 20 and an imaging section 30 . The image inspection apparatus 100 irradiates the object 1 with light L using the illumination unit 20, captures the reflected light R from the object 1 with the imaging unit 30, and uses the information processing unit 10 to perform an image based on the image of the object 1. , determine the state of the object 1 .

The information processing device 10 processes the image captured by the imaging unit 30 using a neural network to determine the state of the object 1 . The information processing device 10 may be configured by a general-purpose computer.

The illumination unit 20 irradiates the object 1 with light. The illumination unit 20 may include one or more lights, and illumination conditions such as illumination direction and illumination intensity may be set in advance.

The photographing unit 30 photographs an image of the object 1 . The imaging unit 30 captures an image of the object 1 illuminated by the illumination unit 20 and transmits the image to the information processing device 10 . The photographing unit 30 may be composed of a general-purpose digital camera, and photographing conditions may be set in advance.

Note that the configuration shown in FIG. 1 is an example, and the image inspection apparatus 100 may include configurations other than the information processing device 10, the illumination unit 20, and the photographing unit 30. The information processing device 10, the illumination unit 20, and A part of the photographing unit 30 may be integrated.

FIG. 2 is a diagram showing functional blocks of the information processing apparatus 10 according to this embodiment. The information processing apparatus 10 includes an acquisition unit 11, a storage unit 12, a determination unit 13, a generation unit 14, and a display unit 10f.

The acquisition unit 11 acquires an image of the object 1 photographed by the photographing unit 30 . Note that the acquisition unit 11 may acquire an image of the target object 1 stored in an external device.

The storage unit 12 stores a first neural network 12a, a second neural network 12b, and a learning image 12c. The first neural network 12a and the second neural network 12b each include at least an input layer, a hidden layer and an output layer each including one or more neurons. The first neural network 12a and the second neural network 12b may be configured by, for example, a CNN (Convolutional Neural Network), but may be other models. Based on the image of the object 1, the first neural network 12a and the second neural network 12b output a numerical value indicating whether the object 1 is normal or abnormal, or output a numerical value indicating the type of the object 1. can be output. The learning image 12c is an image of the object 1 used for learning processing of the first neural network 12a and the second neural network 12b, and may be associated with label data indicating the state of the object 1. The first neural network 12a and the second neural network 12b may be generated, for example, by supervised learning by error backpropagation using the learning image 12c.

Here, compared with the second neural network 12b, the first neural network 12a has a light calculation load, a false negative rate equal to or lower than that of the second neural network 12b, and a large false positive rate. For example, when inspecting whether or not the object 1 has a defect by the image inspection apparatus 100, the case where there is a defect is regarded as positive, and the case where there is no defect is regarded as negative. The rate at which it is determined that there is no defect is the false negative rate, and the rate at which it is determined that there is a defect even though there is actually no defect is the false positive rate. That is, although the first neural network 12a may overdetect defects in comparison with the second neural network 12b, although the computational load is lighter, the rate of overlooking defects is the same as that of the second neural network 12b. It is below. It should be noted that the first neural network 12a and the second neural network 12b may perform ternary output indicating that there is a defect, that there is no defect, or that neither can be determined. In that case, the rate of judging that there is no defect even though there is actually a defect is the false negative rate. The false-positive rate is the rate of false positives.

The determination unit 13 determines the state of the object 1 based on the image of the object 1 using a first neural network 12a and a second neural network 12b each including a plurality of neurons. Here, the determination unit 13 determines the state of the object 1 using the first neural network 12a when the result of determining the state of the object 1 using the first neural network 12a satisfies a predetermined condition. If the result of determining the state of the object 1 using the first neural network 12a does not satisfy a predetermined condition, the determination unit 13 determines the state of the object 1 using the second neural network 12b.

When the first neural network 12a outputs a binary value indicating that the object 1 has a defect or does not have a defect, the predetermined condition may be that the reliability of the determination is equal to or higher than a predetermined value. Further, when the first neural network 12a outputs three values indicating that the object 1 has a defect, does not have a defect, or cannot be determined, the predetermined condition is to output that neither can be determined. good.

In this way, when the judgment result by the first neural network 12a whose computational load is lighter than that of the second neural network 12b satisfies a predetermined condition, the judgment by the second neural network 12b is omitted so that the time required for judgment can be expected. You can shorten the value.

The generating unit 14 generates the second neural network 12b by learning processing using the plurality of learning images 12c, and generates the first neural network 12a by reducing the scale of the second neural network 12b. Here, the process of reducing the scale of the second neural network 12b may be performed by, for example, pruning, distillation or quantization of parameters. In this way, the first neural network 12a smaller in scale than the second neural network 12b can be efficiently generated.

More specifically, of the plurality of neurons included in the hidden layer of the second neural network 12b, the generation unit 14 fixes one or more neurons on the input layer side, and fixes one or more neurons on the output layer side. may be set to zero, and learning processing using a plurality of learning images 12c may be performed to generate the first neural network 12a. As a result, the first neural network 12a can be generated by pruning the hidden layer neurons of the second neural network 12b and re-learning, and the first neural network has a lighter computation load than the second neural network 12b 12a can be produced efficiently.

In addition, the generation unit 14 may generate the first neural network 12a such that the ratio of the scale of the first neural network 12a to the second neural network 12b is equal to or less than a predetermined value. As a result, the memory area occupied by the first neural network 12a can be reduced to a predetermined ratio or less compared to the memory area occupied by the second neural network 12b.

Furthermore, the generation unit 14 generates a plurality of candidate neural networks by reducing the scale of the second neural network 12b by a plurality of different processes, and the plurality of candidate neural networks and the second neural network 12b based on the image. A computation time to determine one state may be measured and a first neural network 12a may be selected from the plurality of candidate neural networks based on the measured computation time. An example of multiple candidate neural networks will be described later with reference to FIG. In this manner, the first neural network 12a can be selected so that the expected value of the time required for determination by the first neural network 12a and the second neural network 12b is the shortest.

The display unit 10f displays the determination result of the determination unit 13 together with the image of the object 1, and displays schematic diagrams of the first neural network 12a and the second neural network 12b generated by the generation unit 14. FIG.

The determination unit 13 temporarily stores the values of predetermined neurons among a plurality of neurons included in the hidden layer of the first neural network 12a, and determines the state of the object 1 using the first neural network 12a. is not satisfied, the state of the object 1 may be determined by the second neural network 12b using the temporarily stored predetermined neuron values. By sharing the value calculated by the first neural network 12a with the second neural network 12b in this way, the time required for the determination by the second neural network 12b can be shortened.

In addition, the generation unit 14 generates the first neural network 12a so that the ratio of a predetermined neuron (a neuron shared with the second neural network 12b) to a plurality of neurons included in the hidden layer of the first neural network 12a is equal to or greater than a predetermined value. A network 12a may be created. In this way, by setting the ratio of neurons of the second neural network 12b that share the values calculated by the first neural network 12a to a predetermined value or more, the time required for determination by the second neural network 12b is shortened by a predetermined time or more. be able to.

FIG. 3 is a diagram showing the physical configuration of the information processing device 10 according to this embodiment. The information processing apparatus 10 includes a CPU (Central Processing Unit) 10a equivalent to a calculation unit, a RAM (Random Access Memory) 10b equivalent to a storage unit, a ROM (Read only memory) 10c equivalent to a storage unit, and a communication unit. 10d, an input unit 10e, and a display unit 10f. These components are connected to each other via a bus so that data can be sent and received. In this example, a case where the information processing apparatus 10 is composed of one computer will be described, but the information processing apparatus 10 may be realized by combining a plurality of computers. Also, the configuration shown in FIG. 3 is an example, and the information processing apparatus 10 may have configurations other than these, or may not have some of these configurations.

The CPU 10a is a control unit that controls the execution of programs stored in the RAM 10b or ROM 10c and performs data calculation and processing. The CPU 10a is an arithmetic unit that executes a program (image inspection program) for determining the state of the object 1 using the first neural network 12a and the second neural network 12b based on the image of the object. The CPU 10a receives various data from the input section 10e and the communication section 10d, displays the calculation results of the data on the display section 10f, and stores them in the RAM 10b and the ROM 10c.

The RAM 10b is a rewritable part of the storage unit, and may be composed of, for example, a semiconductor memory element. The RAM 10b may store data such as the image inspection program executed by the CPU 10a and the learning image 12c. Note that these are examples, and the RAM 10b may store data other than these, or may not store some of them.

The ROM 10c is one of the storage units from which data can be read, and may be composed of, for example, a semiconductor memory element. The ROM 10c may store, for example, an image inspection program and data that is not rewritten.

The communication unit 10d is an interface that connects the information processing device 10 to other devices. The communication unit 10d may be connected to a communication network such as the Internet.

The input unit 10e receives data input from the user, and may include, for example, a keyboard and a touch panel.

The display unit 10f visually displays the calculation result by the CPU 10a, and may be configured by, for example, an LCD (Liquid Crystal Display). The display unit 10f may display an image of the object 1 captured by the imaging unit 30, a learning image 12c, or a determination result.

The image inspection program may be stored in a computer-readable storage medium such as the RAM 10b and the ROM 10c and provided, or may be provided via a communication network connected by the communication unit 10d. In the information processing apparatus 10, the CPU 10a executes the image inspection program to realize various operations described with reference to FIG. It should be noted that these physical configurations are examples, and do not necessarily have to be independent configurations. For example, the information processing apparatus 10 may include an LSI (Large-Scale Integration) in which the CPU 10a, the RAM 10b, and the ROM 10c are integrated.

FIG. 4 is a diagram showing an example of an image IMG captured by the image inspection apparatus 100 according to this embodiment. The image IMG is a captured image of the object 1 . The image IMG includes the first line defect D1 and the second line defect D2. The first line defect D<b>1 is a flaw on the object 1 , and the second line defect D<b>2 is a flaw on the background of the object 1 . The object 1 is abnormal because it includes the first line defect D1, but the second line defect D2 is irrelevant to whether the object 1 is good or bad.

FIG. 5 is a diagram showing an example of the second neural network 12b used by the image inspection apparatus 100 according to this embodiment. The figure schematically shows the input layer IN of the second neural network 12b, the hidden layer HDN, and the neurons OUT1, OUT2, and OUT3 of the output layer. The hidden layer HDN contains multiple neurons.

The second neural network 12b of this example identifies edges included in the image by a plurality of neurons included in the front stage of the hidden layer HDN, and identifies corners included in the image by a plurality of neurons included in the middle stage, and identifies a white line. and black lines, and a specific shape of the object 1 is identified by a plurality of neurons included in the latter stage. These identification contents are simplified for explanation.

Specifically, the first neuron N1 is a neuron that activates when the image contains a right-angled edge, and the second neuron N2 is a neuron that activates when the image contains a horizontal white line. . The second neuron N2 is a neuron that is activated when the image includes a right-angled edge, the fourth neuron N4 is a neuron that is activated when the image includes a downward-sloping white line, and the 5 neuron N5 is a neuron that is activated when the image includes a black line that descends to the right.

Furthermore, the sixth neuron N6 is a neuron activated when the image contains a white line defect, and the seventh neuron N7 is a neuron activated when the image contains a black line defect.

The eighth neuron N8 is a neuron that is activated when the image includes the lower right corner of the object 1, and the ninth neuron N9 is activated when the image includes the upper left corner of the object 1. The 10th neuron N10 is a neuron that is activated when the image contains a hole in the object 1, and the 11th neuron N11 is a neuron that is activated when the image contains the entire object 1. It is a neuron that activates to

The 12th neuron N12, the 13th neuron N13, and the 14th neuron N14 weight and add the outputs of the 6th neuron N6, the 7th neuron N7, and the 11th neuron N11, and calculate the output value by a predetermined nonlinear function, Output to the output layer.

The first neuron OUT1 of the output layer is activated when the object 1 is abnormal (if the object 1 has a defect), and the second neuron OUT2 of the output layer is activated when the object 1 is abnormal or normal. When it cannot be determined whether there is a defect (when it is not possible to determine whether the object 1 has a defect or not), the third neuron OUT3 in the output layer is activated when the object 1 is normal (when the object 1 has no defect). ).

FIG. 6 is a diagram showing an example of the first neural network 12a used by the image inspection apparatus 100 according to this embodiment. The first neural network 12a of this example is generated by pruning a plurality of neurons included in the hidden layer HDN of the second neural network 12b.

Compared to the second neural network 12b, the first neural network 12a has a first neuron N1, a second neuron N2, a third neuron N3, a fourth neuron N4, a sixth neuron N6, an eighth neuron N8, and a ninth neuron. N9, 10th neuron N10, 11th neuron N11, 12th neuron N12 and 14th neuron N14 are pruned to reduce their scale.

The generation unit 14 fixes one or more neurons on the input layer side among the plurality of neurons included in the hidden layer HDN of the second neural network 12b, and fixes one or more neurons on the output layer side (in this example, 1st neuron N1, 2nd neuron N2, 3rd neuron N3, 4th neuron N4, 6th neuron N6, 8th neuron N8, 9th neuron N9, 10th neuron N10, 11th neuron N11, 12th neuron N12 and The first neural network 12a may be generated by performing learning processing using a plurality of learning images 12c with the weight parameter and bias parameter of the 14 neurons N14) set to zero.

Further, the determination unit 13 temporarily stores the value of a predetermined neuron S among a plurality of neurons included in the hidden layer HDN of the first neural network 12a, and determines the state of the object 1 using the first neural network 12a. If the result does not satisfy the predetermined condition, the state of the object 1 may be determined by the second neural network 12b using the temporarily stored predetermined neuron S value. By sharing the value calculated by the predetermined neuron S of the first neural network 12a with the second neural network 12b, the time required for determination by the second neural network 12b can be shortened.

FIG. 7 is a diagram showing an example of a plurality of candidate neural networks generated by the image inspection apparatus 100 according to this embodiment. The generation unit 14 generates candidate neural networks such that the ratio of scale to the second neural network 12b is equal to or less than a predetermined value, or generates candidate neural networks such that the ratio of neurons shared with the second neural network 12b is equal to or greater than a predetermined value. generate a network. The generation unit 14 allows the false positive rate of the candidate neural network to be greater than that of the second neural network 12b, while generating the candidate neural network so that the false negative rate of the candidate neural network is equal to or lower than that of the second neural network 12b. You may generate a network.

In this example, the first candidate neural network has a scale ratio of 40% with respect to the second neural network 12b, a ratio of neurons shared with the second neural network 12b of 10%, and the object 1 is normal. 4% of the cases cannot be judged whether it is abnormal or not. In addition, the second candidate neural network has a scale ratio of 40% with respect to the second neural network 12b, a ratio of neurons shared with the second neural network 12b is 20%, and the object 1 is normal or abnormal. 6% are unable to determine whether or not In addition, the third candidate neural network has a scale ratio of 20% with respect to the second neural network 12b, a ratio of neurons shared with the second neural network 12b is 5%, and the object 1 is normal or abnormal. 8% are unable to determine whether or not The fourth candidate neural network has a scale ratio of 30% to the second neural network 12b, a neuron ratio of 15% shared with the second neural network 12b, and the object 1 is normal or abnormal. 7% are unable to determine either

For example, if it is desired to reduce the rate of failure to determine whether the object 1 is normal or abnormal, the first candidate neural network may be selected as the first neural network 12a. Also, if it is desired to increase the ratio of neurons shared with the second neural network 12b, the second candidate neural network may be selected as the first neural network 12a. On the other hand, if it is desired to keep the scale of the network as small as possible, the third candidate neural network may be selected as the first neural network 12a. Also, if it is desired to generate a network having all items in a well-balanced manner, the fourth candidate neural network may be selected as the first neural network 12a.

Further, the generation unit 14 selects the first neural network 12a from the plurality of candidate neural networks based on the calculation time for determining the state of the object 1 based on the image by the plurality of candidate neural networks and the second neural network 12b. You can

FIG. 8 is a flowchart of the first neural network selection process executed by the image inspection apparatus 100 according to this embodiment. First, the image inspection apparatus 100 generates the second neural network 12b by learning processing using the learning image 12c (S10).

After that, the image inspection apparatus 100 performs N scale reductions on the second neural network 12b (S11), and by learning processing using the learning image 12c, the first candidate to the Nth candidate of the first neural network. A candidate is generated (S12). Here, N is a natural number of 2 or more.

Then, the image inspection apparatus 100 measures the calculation time of the first to N-th candidates of the first neural network and the calculation time of the second neural network 12b (S13), and based on the calculation time, the first candidate to The first neural network 12a is selected from among the Nth candidates (S14). The above completes the first neural network selection process.

FIG. 9 is a flowchart of state determination processing executed by the image inspection apparatus 100 according to this embodiment. First, the image inspection apparatus 100 takes an image of the object 1 (S20).

After that, the image inspection apparatus 100 inputs the image to the first neural network 12a and calculates the active state of the output layer (S21). Then, it is determined whether the value of the neuron in the output layer corresponding to OK (object 1 is normal) or NG (object 1 is above) is maximum (S22). That is, it is determined whether the value of the neuron in the output layer corresponding to the gray state in which the object 1 cannot be determined to be normal or abnormal is the maximum.

If the neuron value corresponding to OK or NG is the maximum (S22: YES), the image inspection apparatus 100 determines the state of the object 1 based on the output of the first neural network 12a (S23).

On the other hand, if the neuron value corresponding to OK or NG is not the maximum (S22: NO), that is, if the neuron value corresponding to gray is maximum, the image inspection apparatus 100 inputs the image to the second neural network 12b. Then, the active state of the output layer is calculated (S24), and the state of the object 1 is determined based on the output of the second neural network 12b (S25). With the above, the determination processing ends.

As described above, according to the image inspection apparatus 100 according to the present embodiment, when the determination result by the first neural network 12a whose computational load is lighter than that of the second neural network 12b satisfies a predetermined condition, the second neural network 12b is used. By omitting the determination, the expected value of the time required for the determination can be shortened.

The embodiments described above are for facilitating understanding of the present invention, and are not intended to limit and interpret the present invention. Each element included in the embodiment and its arrangement, materials, conditions, shape, size, etc. are not limited to those illustrated and can be changed as appropriate. Also, it is possible to partially replace or combine the configurations shown in different embodiments.

Aspects in this embodiment include the following disclosures.

[Appendix 1]
a photographing unit (30) for photographing an image of an object (1);
a determination unit (13) that determines the state of the object (1) based on the image using a first neural network (12a) and a second neural network (12b) each including a plurality of neurons; ,
Compared to the second neural network (12b), the first neural network (12a) has a light computational load, an equal or lower false negative rate, and a large false positive rate,
The determination unit (13)
When the result of determining the state of the object (1) using the first neural network (12a) satisfies a predetermined condition, the first neural network (12a) is used to determine the state of the object (1). to determine
If the result of determining the state of the object (1) using the first neural network (12a) does not satisfy the predetermined condition, the object (1) using the second neural network (12b) determine the state of
An image inspection device (100).

DESCRIPTION OF SYMBOLS 10... Information processing apparatus 10a...CPU, 10b...RAM, 10c...ROM, 10d...Communication part, 10e...Input part, 10f...Display part, 11... Acquisition part, 12... Storage part, 12a...First neural network, 12b... Second neural network 12c... Learning image 13... Judging unit 14... Generating unit 20... Illuminating unit 30... Imaging unit 100... Image inspection device

Claims (5)

a photographing unit for photographing an image of an object;
a determination unit that determines the state of the object based on the image using a first neural network and a second neural network each including a plurality of neurons;
a generation unit that generates the second neural network by learning processing using a plurality of learning images and generates the first neural network by processing for reducing the scale of the second neural network;
with
Compared to the second neural network, the first neural network has a light computational load, an equal or lower false negative rate, and a large false positive rate,
The determination unit is
If the result of determining the state of the object using the first neural network satisfies a predetermined condition, determining the state of the object using the first neural network;
if the result of determining the state of the object using the first neural network does not satisfy the predetermined condition, determining the state of the object using the second neural network;
The generation unit generates a plurality of candidate neural networks by reducing the scale of the second neural network by a plurality of different processes, and the target object based on the image by the plurality of candidate neural networks and the second neural network. measuring the computation time to determine the state of, and selecting the first neural network from the plurality of candidate neural networks based on the computation time;
Image inspection equipment. The generator fixes one or more neurons on the input layer side among a plurality of neurons included in the hidden layer of the second neural network, and sets the parameters of one or more neurons on the output layer side to zero, generating the first neural network by performing a learning process using the plurality of learning images;
The image inspection apparatus according to claim 1 . The generation unit generates the first neural network such that the ratio of the scale of the first neural network to the second neural network is equal to or less than a predetermined value.
The image inspection apparatus according to claim 1 or 2 . The determination unit temporarily stores values of predetermined neurons among a plurality of neurons included in a hidden layer of the first neural network, and determines the state of the object using the first neural network. If a predetermined condition is not satisfied, the second neural network determines the state of the object using the temporarily stored predetermined neuron value.
The image inspection apparatus according to any one of claims 1 to 3 . The generation unit generates the first neural network such that a ratio of the predetermined neuron to the plurality of neurons included in the hidden layer of the first neural network is equal to or greater than a predetermined value.
The image inspection apparatus according to claim 4 .

Images