A processing method and a processing device using the same
By summarizing electrical symbols of similar shapes into systems, using neural networks for learning and detection, extracting and clustering of intermediate layer data, learning difficulties and accuracy reduction problems caused by multiple types of electrical symbols are solved, and more efficient electrical symbols are achieved extract.
Patent Information
- Application Number
- CN201980021870.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-03-26
- Filing Date
- 2019-03-25
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2039-03-25
AI Technical Summary
In architectural design drawings, various types of electrical symbols exist, making it difficult for neural networks to learn and resulting in reduced processing accuracy.
Electrical symbols of similar shapes are summarized into systems, learn and detect through neural networks, extract specific intermediate layer data and cluster them, reducing the number of pattern types to reduce the learning workload.
It effectively suppresses the increase in learning workload, improves processing accuracy, and reduces the possibility of misjudgment.
Smart Images

Figure CN111902827B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to processing technology, and more particularly to a processing method for performing processing on an image and a processing device using the same. Background Art
[0002] In recent years, construction drawings can be easily prepared using CAD (Computer-Aided Design) systems and the like. However, in the data of construction drawings, there may be situations where there is no compatibility due to different systems used for preparation, or where the data cannot be used due to the passage of time or changes in practitioners. Therefore, it takes a lot of effort to extract electrical symbols from the already prepared construction drawings. In order to extract the electrical symbols recorded in the construction drawings, the following processing is performed: contour data of the image is prepared based on the running data of the image image data obtained by reading the image of the construction drawing, and contour data whose circularity of the contour ring corresponding to the contour data is above a threshold value is regarded as an electrical symbol (for example, refer to Patent Document 1).
[0003] [Prior art literature]
[0004] [Patent Document]
[0005] Patent Document 1: Japanese Patent Application Laid-Open No. 10-111937 Summary of the invention
[0006] [Problems to be solved by the invention]
[0007] In order to extract patterns such as electrical symbols from images such as architectural design drawings, a neural network can be used. As long as the pattern can be extracted, multiple patterns contained in the image can be accumulated according to the type of pattern. However, since there are many types of patterns, it is difficult for the neural network to learn all patterns. On the other hand, when the learning is insufficient, the accuracy of the processing will be reduced.
[0008] The present disclosure has been made in view of such a situation, and an object of the present disclosure is to provide a technology that suppresses an increase in learning workload and a decrease in processing accuracy.
[0009] [Technical solutions for solving technical problems]
[0010] In order to solve the above-mentioned problem, a processing device of a scheme of the present invention is a processing device that accumulates multiple patterns contained in an image according to the type of pattern, and includes: an input unit that inputs an image; a first processing unit that performs processing of a neural network that uses a system that classifies the types of patterns and has a smaller number of types than the number of pattern types as supervisory data to learn the image input to the input unit, thereby detecting the multiple patterns contained in the image as a certain system; an extraction unit that extracts parts contained in the multiple systems from the image based on the positions of each system detected in the first processing unit; a second processing unit that obtains intermediate data of a specific intermediate layer corresponding to each system by performing processing of a neural network that uses a system that classifies the types of patterns and has a smaller number of types than the number of pattern types as supervisory data to learn the multiple parts extracted in the extraction unit; a clustering unit that performs clustering corresponding to the number of pattern types on the intermediate data obtained in the second processing unit; and a calculation unit that calculates the result of clustering in the clustering unit, that is, the number of patterns contained in each cluster.
[0011] Another scheme of the present disclosure is a processing method. The method is a processing method for accumulating multiple patterns contained in an image according to the types of patterns, including the steps of inputting an image, performing a neural network process on the input image to learn a system that classifies the types of patterns and has a smaller number of types than the number of pattern types as supervisory data, thereby detecting the multiple patterns contained in the image as a certain system, extracting parts contained in the multiple systems from the image based on the positions of the detected systems, performing a neural network process on the extracted multiple parts to learn a system that classifies the types of patterns and has a smaller number of types than the number of pattern types as supervisory data, thereby obtaining intermediate data of a specific intermediate layer corresponding to each system, performing clustering corresponding to the number of pattern types on the obtained intermediate data, and calculating the result of clustering, that is, the number of patterns contained in each cluster.
[0012] In addition, any combination of the above-mentioned constituent elements and any expression of the present disclosure in the form of a method, an apparatus, a system, a computer program, or a recording medium recording the computer program may also be effective as aspects of the present disclosure.
[0013] [Effects of the invention]
[0014] According to the present disclosure, it is possible to suppress both an increase in learning workload and a decrease in processing accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 FIG. 1 is a diagram showing an image to be processed in the embodiment.
[0016] Figure 2 Yes means Figure 1 The images may contain patterns and systematic diagrams.
[0017] Figure 3 (a) - Figure 3 (b) is a diagram showing the structure of the processing device of the embodiment.
[0018] Figure 4 Yes means Figure 3 (a) - Figure 3 (b) is a diagram showing an overview of the processing in the processing unit.
[0019] Figure 5 It means that it is input into Figure 3 (a) is a diagram of the data structure of the supervision data of the first input part.
[0020] Figure 6 (a) - Figure 6 (b) means Figure 3 (b) is a diagram of the processing result of the processing device.
[0021] Figure 7 Yes means Figure 3 (b) is a diagram showing an overview of the processing in the clustering unit.
[0022] Figure 8 Yes means Figure 3 (b) is a flowchart of the processing steps in the processing device. DETAILED DESCRIPTION
[0023] Before describing the embodiments of the present disclosure in detail, an overview of the embodiments is described. The embodiments relate to a processing device that accumulates the number of electrical symbols by type based on a construction drawing showing a plurality of electrical symbols. In order to improve the accuracy of extracting electrical symbols from construction drawings, it is advantageous to use a neural network. However, there are more than dozens of types of electrical symbols, and there are also electrical symbols of similar shapes. Therefore, it is difficult to make the neural network learn for all electrical symbols, and in the case of the appearance of new electrical symbols, there is a possibility of misjudgment. Furthermore, there is also a possibility of misjudgment when learning is insufficient.
[0024] In order to cope with this situation, in this embodiment, the following method is used: by classifying electrical symbols of similar shapes into "groups", multiple patterns are classified into groups with a smaller number of types than the number of electrical symbols. Hereinafter, electrical symbols are referred to as "patterns", groups are referred to as "systems", and construction drawings are referred to as "images". Here, for example, patterns are classified into three types of systems: intercom systems, differential smoke sensor systems, and photoelectric smoke sensor systems. Under such circumstances, the neural network in the processing device learns for these systems. When an image is input to the processing device, each system shown in the image is detected by the processing of the neural network, and the position of each system is determined. The processing device extracts a portion containing each system from the image based on the position of each system. In addition, the processing device outputs intermediate data of a specific intermediate layer by performing neural network processing on each extracted portion. Furthermore, after performing clustering on the intermediate data, the processing device calculates the number of patterns contained in each cluster.
[0025] Figure 1 The image 10 to be processed is represented. As mentioned above, an example of the image 10 is a construction drawing. The image 10 includes the first pattern 20a to the 26th pattern 20z, the 27th pattern 20aa, the 28th pattern 20ab, the 29th pattern 20ac, and the 30th pattern 20ad, which are collectively referred to as patterns 20. Each pattern 20 is an electrical symbol. Here, the pattern 20 included in the image 10 is set to "30", but it is not limited to this. Because when the shape of the pattern 20 is different, it represents other electrical symbols, it is equivalent to other machines. For example, the fifth pattern 20e and the sixth pattern 20f are the same shape, but the fifth pattern 20e and the first pattern 20a are different shapes. In the case of making a quotation based on such an image 10, the user needs to count the number of patterns 20 according to the type of pattern 20. As the size of the image 10 increases and the number of patterns 20 included in the image 10 increases, the work required for this work will increase.
[0026] Figure 2Represents patterns 20 and systems that may be included in the image 10. For similar machines, patterns 20 of similar shapes are determined. Here, patterns 20 of similar shapes are summarized as one system. As shown in the figure, three types of systems are specified, namely, an intercom system, a differential smoke sensor system, and a photoelectric smoke sensor system. In addition, the intercom system includes the first intercom 30a to the fourth intercom 30d collectively referred to as the intercom 30. The differential smoke sensor system includes the first differential smoke sensor 32a to the sixth differential smoke sensor 32f collectively referred to as the differential smoke sensor 32. The photoelectric smoke sensor system includes the first photoelectric smoke sensor 34a to the sixth photoelectric smoke sensor 34f collectively referred to as the photoelectric smoke sensor 34. The intercom 30, the differential smoke sensor 32, and the photoelectric smoke sensor 34 are also one of the patterns 20. For example, Figure 1 The first pattern 20a is the second intercom 30b, and the second pattern 20b is the sixth photoelectric smoke sensor 34f. In this way, the system can be said to classify the types of patterns 20. In addition, the number of types of the system is less than the number of types of patterns 20.
[0027] Figure 3 (a) - Figure 3 (b) shows the structure of the processing device 100. In particular, Figure 3 (a) represents the structure used for learning processing, Figure 3 (b) shows a configuration for accumulation processing. The accumulation processing is a process of accumulating a plurality of patterns 20 included in the image 10 by type while systematically classifying the patterns 20 using a neural network for the image 10 . Figure 3 The processing device 100 in (a) and Figure 3 The processing devices 100 in (b) may be the same device or different devices.
[0028] The processing device 100 includes a first input unit 110, a second input unit 112, and a processing unit 114 as a configuration for learning processing, and includes an input unit 120, a processing unit 114, an extraction unit 124, a clustering unit 128, a calculation unit 130, and an output unit 132 as a configuration for accumulation processing. In addition, the processing unit 114 includes a first processing unit 122 and a second processing unit 126. Here, in the learning processing, the processing unit 114 is learned, and in the accumulation processing, the processing unit 114 is used. Before describing the configuration of the processing device 100, the configuration of the processing unit 114 will be described.
[0029] Figure 4The processing in the processing unit 114 is summarized. The processing unit 114 includes a first convolution layer 142a, a second convolution layer 142b, a third convolution layer 142c, a fourth convolution layer 142d, a fifth convolution layer 142e, and a sixth convolution layer 142f, which are collectively referred to as convolution layers 142, and a first pooling layer 144a, a second pooling layer 144b, a third pooling layer 144c, a fourth pooling layer 144d, and a fifth pooling layer 144e, which are collectively referred to as pooling layers 144. Such convolution layers 142 and pooling layers 144 represent a block diagram for executing each processing. The input image 140 is an image to be processed in the processing device 100. Since the neural network in the processing unit 114 does not include a fully connected layer, no restriction on the size of the input image 140 is set. Therefore, the input image 140 may also be the aforementioned image 10. The input image 140 is input to the first convolutional layer 142a.
[0030] Each convolution layer 142 is represented as a right parallelepiped. The size of the square surface having the depth direction and the height direction of the convolution layer 142 represents the size of the spatial dimension of the image processed in the convolution layer 142, that is, the size of the image. In the convolution layer 142, for the image, spatial filters whose sizes are smaller than the image size are staggered to perform spatial filtering. Because spatial filtering is a well-known technology, the description is omitted here. This spatial filtering is equivalent to convolution processing, and the feature quantity of the image is extracted through convolution processing. It is also possible that padding, etc. are performed in the convolution layer 142. Furthermore, it is also possible that in the convolution layer 142, for the image, multiple spatial filters are used in parallel, thereby performing multiple spatial filtering in parallel. By using multiple spatial filters in parallel, the image will increase. The number of spatial filters used in parallel in the convolution layer 142 is called the number of channels, which is represented by the length of the convolution layer 142 in the left and right directions.
[0031] Each pooling layer 144 is also constructed in the same manner as the convolution layer 142. In the pooling layer 144, the size of the image is reduced by grouping multiple pixels contained in an arbitrary area in the image into one pixel. Here, in order to group multiple pixels into one pixel, average pooling or maximum pooling is performed. In average pooling, the average value of multiple pixel values in the area is used for one pixel, and in maximum pooling, the maximum value of multiple pixel values in the area is used for one pixel. Pooling is performed to enhance robustness against translational motion of a representative value or an average value in the region of interest.
[0032] Here, the first convolution layer 142a, the first pooling layer 144a, the second convolution layer 142b, the second pooling layer 144b, the third convolution layer 142c, the third pooling layer 144c, the fourth convolution layer 142d, the fourth pooling layer 144d, the fifth convolution layer 142e, the fifth pooling layer 144e, and the sixth convolution layer 142f are sequentially arranged. That is, the convolution process and the pooling process are repeatedly performed. In addition, by repeatedly performing the convolution process and the pooling process, the size of the image is successively reduced, and the sixth pooling layer 144f outputs an image with more than one channel number (hereinafter referred to as "output image"). Here, as an example, the channel of the output image is set to "3".
[0033] Based on the configuration of the processing unit 114, Figure 3 The learning process in the processing device 100 is described in (a). The first input unit 110 receives a known output image, i.e., supervisory data, and the second input unit 112 receives a learning image corresponding to the supervisory data received in the first input unit 110. The processing unit 114 has Figure 4 The convolutional layer 142 is configured to learn the coefficients of the spatial filters of each convolutional layer 142 based on the supervision data received in the first input unit 110 and the learning image received in the second input unit 112.
[0034] The learning image includes a pattern 20 of any one of an intercom 30, a differential smoke sensor 32, and a photoelectric smoke sensor 34. The supervisory data is an intercom system when the intercom 30 is included in the learning image, a differential smoke sensor system when the differential smoke sensor 32 is included as the learning image, and a photoelectric smoke sensor system when the photoelectric smoke sensor 34 is included as the learning image. Figure 5 The data structure of the supervisory data input to the first input unit 110 is shown. The supervisory data has three channels, and any one of the intercom system, the differential smoke sensor system, and the photoelectric smoke sensor system is represented by the learning image.
[0035] That is, the size of the supervisory data has a spatial dimension of 1×1. Therefore, the supervisory data does not have the position information of the pattern 20 contained in the learning image, and the supervisory data corresponding to one channel only represents any one of the intercom system, the differential smoke sensor system, and the photoelectric smoke sensor system. On the other hand, the learning image is an original image that will output the supervisory data when the accumulation processing is accurately performed, and its size is determined in such a way that the supervisory data has a spatial dimension of 1×1. Here, since there is no restriction on the size of the input image 140, the learning image and the input image 140 can also be of different sizes. Since the learning of the coefficients of the spatial filter under the condition of using such supervisory data and learning images can use known techniques, the description is omitted here.
[0036] Next, use Figure 3 (b) of the present invention will explain the accumulation processing in the processing device 100. Figure 3 (a) and Figure 3 In the case where the processing device 100 in (b) is configured as a different device, according to Figure 3 The coefficients of the spatial filter derived by the learning of the processing unit 114 in (a) are set to Figure 3 The processing unit 114 in (b).
[0037] The input unit 120 receives the image 10. Here, the size of the image 10 is set larger than the size of the learning image. The input unit 120 outputs the image 10 to the first processing unit 122. The first processing unit 122 receives the image 10 from the input unit 120. The first processing unit 122 has Figure 4 In the structure shown, the received image 10 is equivalent to the input image 140. Therefore, hereinafter, the image 10 is referred to as the input image 140. The first processing unit 122 performs convolutional neural network processing except for the fully connected layer on the input image 140. In the convolutional neural network in the first processing unit 122, the spatial filter of the convolutional layer 142 is learned for the supervisory data for the system whose number of types is less than the number of types of the pattern 20. In particular, the learning of the spatial filter of the convolutional layer 142 is performed for the system corresponding to the three channels one by one. As a result, the first processing unit 122 detects each of the multiple patterns 20 contained in the image 10 as a certain system. As mentioned above, the size of the image 10 is larger than the size of the learning image, so the first processing unit 122 also detects the position of the system. Figure 6 (a) - Figure 6 (b) shows the processing result of the processing device 100. Figure 6 (a) indicates that Figure 1 The number of each system detected when the processing of the first processing unit 122 is performed on the image 10. Figure 6(b) will be described later. Next, return to Figure 3 The first processing unit 122 outputs the detected plurality of combinations of systems and positions to the extraction unit 124 .
[0038] The extraction unit 124 receives a plurality of combinations from the first processing unit 122. The extraction unit 124 obtains the detected system and position from each combination, and extracts the portion included in the pattern 20 corresponding to the system from the image 10. The portion has an arbitrary shape and an arbitrary size centered on the pattern 20 corresponding to the system. The size of the portion is smaller than the size of the image 10. In addition, the size of the portion may also be larger than the size of the learning image. Through this processing, the extraction unit 124 extracts a plurality of portions. The extraction unit 124 outputs the plurality of portions to the second processing unit 126.
[0039] The second processing unit 126 shares the neural network with the first processing unit 122, and thus uses the neural network used in the first processing unit 122. The second processing unit 126 performs neural network processing on the plurality of parts extracted by the extraction unit 124. The neural network processing is the same as that in the first processing unit 122, and thus the description thereof is omitted here. Figure 4 In this way, the intermediate data 148 of the specific intermediate layer corresponding to each part is obtained. The intermediate data 148 is data output from the intermediate layer one or two before the output of the second processing unit 126, and is represented as n-dimensional data such as (v1, v2, . . . , vn). The second processing unit 126 outputs the intermediate data 148 to the clustering unit 128 for each part, that is, for each detected system.
[0040] The clustering unit 128 performs clustering corresponding to the number of types of the patterns 20 on the plurality of intermediate data 148 obtained in the second processing unit 126. Here, for clustering, for example, the K-means method, self-organizing map, minimum mean dispersion method, NN (Nearest Neighbor) method, shortest distance method, K-NN (K Nearest Neighbor) method, and Ward method are used. These are well-known techniques, and thus description thereof is omitted here.
[0041] Figure 7The following is an overview of the processing in the clustering unit 128. Here, for the sake of clarity, it is assumed that the intermediate data 148 is two-dimensional data (v1, v2). The clustering unit 128 configures the plurality of intermediate data 148 in an orthogonal coordinate system with v1 and v2 as axes. The clustering unit 128 generates the first cluster 150a to the eleventh cluster 150k, which are collectively referred to as clusters 150, by performing clustering on the plurality of intermediate data 148 configured in the orthogonal coordinate system. Each cluster 150 corresponds to any one of the patterns 20. For example, the first cluster 150a corresponds to the second intercom 30b. Next, returning to Figure 3 (b).
[0042] The calculation unit 130 calculates the result of clustering by the clustering unit 128 , that is, the number of patterns 20 included in each cluster. Figure 6 (b) shows the calculation result of the calculation unit 130. The second intercom 30b is calculated as 1, and the second differential smoke sensor 32b is calculated as 8. Figure 3 The output unit 132 outputs the calculation result obtained by the calculation unit 130 .
[0043] The main body of the device, system or method in the present disclosure includes a computer. The computer implements the functions of the main body of the device, system or method in the present disclosure by executing a program. The computer includes a processor that acts according to the program as the main hardware structure. As long as the processor can implement the function by executing the program, its type is not particularly limited. The processor is composed of one or more electronic circuits including a semiconductor integrated circuit (IC) or LSI (Large Scale Integration). Multiple electronic circuits can be integrated into one chip or provided in multiple chips. Multiple chips can be integrated into one device or provided in multiple devices. The program is recorded in a non-temporary recording medium such as a computer-readable ROM, an optical disk, and a hard disk drive. The program can be pre-contained in the recording medium or supplied to the recording medium via a wide area communication network including the Internet.
[0044] The operation of the processing device 100 having the above configuration will be described below. Figure 8 1 is a flowchart showing the processing steps in the processing device 100. The input unit 120 inputs the image 10 (S10). The first processing unit 122 detects the system (S12). The extraction unit 124 extracts the part including the system (S14). The second processing unit 126 obtains the intermediate data 148 of the intermediate layer (S16). The clustering unit 128 performs clustering (S18). The calculation unit 130 calculates the number of patterns 20 (S20).
[0045] According to the present embodiment, since the neural network processing is performed on the part including the system detected from the image 10, the influence of the pattern other than the pattern 20 included in the image 10 can be reduced. In addition, since the influence of the pattern other than the pattern 20 included in the image 10 is reduced, the accuracy of the processing can be improved. Since the intermediate data 148 in the processing of the neural network is obtained, the amount of information can be increased. In addition, since the amount of information is increased, the accuracy of the processing can be improved. In addition, since the number of patterns 20 included in each cluster 150 is calculated by performing clustering on the intermediate data 148, it is possible to suppress both the increase in learning workload and the reduction in processing accuracy.
[0046] In addition, since the intermediate data 148 of the intermediate layer before one or two outputs is obtained, data with higher accuracy and more information can be obtained. In addition, since the second processing unit 126 uses the neural network used in the first processing unit 122, the learning results can be shared. In addition, since a convolutional neural network that has learned the spatial filter of the convolution layer 142 is used for supervisory data with a spatial dimension of 1×1, the workload of producing supervisory data can be reduced. In addition, since the workload of producing supervisory data is reduced, the increase in learning workload can be suppressed. In addition, since the workload of producing supervisory data is reduced, the amount of supervisory data can be increased. In addition, since the amount of supervisory data is increased, the accuracy of learning can be improved.
[0047] The outline of one scheme of the present disclosure is as follows. A processing device (100) of one scheme of the present disclosure is a processing device (100) for accumulating a plurality of patterns contained in an image (10) by pattern type, comprising: an input unit (120) that inputs an image (10); a first processing unit (122) that performs a neural network process for learning a system that classifies the types of patterns (20) and has a smaller number of types than the number of types of patterns (20) as supervisory data for the image (10) input to the input unit (120), thereby detecting each of the plurality of patterns (20) contained in the image (10) as a certain system; and an extraction unit (124) that extracts the positions of the respective systems from the image (10) based on the positions of the respective systems detected in the first processing unit (122). 0) to extract parts included in multiple systems; a second processing unit (126) that processes a neural network for learning a system that classifies types of patterns (20) and has a smaller number of types than the number of types of patterns (20) as supervisory data for the multiple parts extracted in the extraction unit (124), thereby obtaining intermediate data (148) of a specific intermediate layer corresponding to each system; a clustering unit (128) that performs clustering corresponding to the number of types of patterns (20) on the intermediate data (148) obtained in the second processing unit (126); and a calculation unit (130) that calculates the result of clustering in the clustering unit (128), i.e., the number of patterns (20) included in each cluster (150).
[0048] The second processing unit (126) may obtain the intermediate data (148) of the intermediate layer one or two layers before the output.
[0049] The second processing unit (126) may use the neural network used in the first processing unit (122).
[0050] Alternatively, the neural network used in the first processing unit (122) and the second processing unit (126) includes a convolutional layer (142) and a pooling layer (144), which is a convolutional neural network excluding a fully connected layer, in which the filter of the convolutional layer (142) is learned for a processing result having a spatial dimension of 1×1.
[0051] Another embodiment of the present disclosure is a processing method. The method is a processing method for accumulating multiple patterns (20) contained in an image (10) according to the types of the patterns (20), comprising: a step of inputting an image (10), a step of performing a neural network process for learning a system that classifies the types of patterns (20) and has a smaller number of types than the number of types of patterns (20) as supervisory data for the input image (10), thereby detecting the multiple patterns (20) contained in the image (10) as a certain system, and a step of accumulating the multiple patterns (20) contained in the image (10) according to the positions of the detected systems. The invention further comprises the steps of extracting a plurality of parts including a system, executing a neural network processing for the extracted plurality of parts, which uses a system for classifying types of patterns (20) and having a smaller number of types than the number of types of patterns (20) as supervisory data to make the system learn, thereby obtaining intermediate data (148) of a specific intermediate layer corresponding to each system, executing clustering corresponding to the number of types of patterns (20) for the obtained intermediate data (148), and calculating the result of clustering, i.e., the number of patterns (20) included in each cluster (150).
[0052] The present disclosure has been described above based on the embodiments. It should be understood by those skilled in the art that the embodiments are merely illustrative, and various modifications may exist in the combination of their constituent elements or processing procedures, and such modifications are also within the scope of the present disclosure.
[0053] In the configuration of the processing unit 114 in this embodiment, a plurality of convolutional layers 142 and a plurality of pooling layers 144 are alternately arranged. However, this is not limiting, and for example, the processing unit 114 may have a configuration such as a GoogLeNet system or a DenseNet system. According to this variation, the degree of freedom of configuration can be increased.
[0054] The processing unit 114 in this embodiment uses a convolutional neural network other than a fully connected layer, that is, a fully convolutional neural network. However, it is not limited to this, and for example, a convolutional neural network including a fully connected layer may also be used. In this case, since the size of the image to be input to the input unit 120 is fixed, the image 10 is divided into fixed sizes and input to the input unit 120. According to this modification, the degree of freedom of the configuration can be increased.
[0055] [Explanation of Reference Numerals]
[0056] 10 image, 20 pattern, 30 intercom, 32 differential smoke sensor, 34 photoelectric smoke sensor, 100 processing device, 110 first input unit, 112 second input unit, 114 processing unit, 120 input unit, 122 first processing unit, 124 extraction unit, 126 second processing unit, 128 clustering unit, 130 calculation unit, 132 output unit, 140 input image, 142 convolution layer, 144 pooling layer.
[0057] [Industrial Applicability]
[0058] According to the present disclosure, it is possible to suppress both an increase in learning workload and a decrease in processing accuracy.
Claims
1. A processing device that accumulates a plurality of electrical symbols included in a construction image according to the type of the electrical symbols; The processing device is characterized by comprising: An input unit that inputs a construction image, The first processing unit detects each of the plurality of electrical symbols included in the construction image as a certain system by executing a neural network processing for learning a system that classifies the types of electrical symbols and has a smaller number of types than the number of electrical symbols as supervisory data for the construction image input to the input unit. an extracting unit that extracts a plurality of parts including the systems from the construction image based on the positions of the systems detected by the first processing unit, a second processing unit that uses the neural network used in the first processing unit to obtain intermediate data of a specific intermediate layer corresponding to each system for the plurality of parts extracted in the extraction unit, a clustering unit that performs clustering on the intermediate data obtained in the second processing unit, and The calculation unit counts the number of electrical symbols included in each cluster as a result of the clustering performed by the clustering unit.
2. The processing device according to claim 1, characterized in that The second processing unit obtains the intermediate data of the intermediate layer one or two layers before being output.
3. The processing device according to claim 1, characterized in that The neural network used in the above-mentioned first processing unit and the above-mentioned second processing unit includes a convolution layer and a pooling layer, which is a convolutional neural network except for the fully connected layer. In this convolutional neural network, the filter of the above-mentioned convolution layer is learned for the processing result with a spatial dimension of 1×1.
4. A processing method, which accumulates a plurality of electrical symbols contained in a construction image according to the types of electrical symbols; The processing method is characterized by comprising: Steps for importing construction images, A step of detecting a plurality of electrical symbols contained in the construction image as a certain system by executing a neural network processing for learning by using a system that classifies the types of electrical symbols and has a smaller number of types than the number of electrical symbols as supervisory data for input construction images, a step of extracting a plurality of parts including the systems from the construction image based on the detected positions of the systems, A step of obtaining intermediate data of a specific intermediate layer corresponding to each system using the neural network used in the above detection step for the extracted multiple parts, Performing clustering steps on the obtained intermediate data, and The step of counting the clustering results, that is, the number of electrical symbols contained in each cluster.
