Drawing analysis system

Abstract

This system provides drawing analysis capabilities that enable deep learning-based object detection, transfer learning, mapping, cost calculation, and data output. [Solution] The drawing analysis system includes the following configuration: A dataset generation unit 1 that generates multiple types of graphic datasets including equipment symbols. A learning unit 2 that learns the features of equipment symbols based on the dataset using a deep learning model. A detection unit 3 that applies the deep learning model to unknown dataset data and automatically detects and classifies multiple types of equipment symbols. Although not essential to the present invention, it is also possible to include a transfer learning unit 4 that enables feature extraction specific to equipment symbols in drawings through transfer learning.

Description

Technical Field

[0001] The present invention relates to a drawing analysis system used for detecting facility symbols included in drawings, particularly lighting fixture symbols.

Background Art

[0002] As shown in Patent Document 1, there has conventionally been known an invention that reads lighting symbols in architectural drawings using image recognition technology and presents to a user the cost and lease fee when replacing fluorescent lamps and incandescent bulbs indicated by the lighting symbols with LED lighting.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, this prior art stores a plurality of types of lighting symbols in a database in advance, compares the image read from the drawing with the lighting symbols in the database, and determines the appliance represented by the lighting symbol. Therefore, there are limitations in the comparison and collation between the lighting symbols stored in the database and the actually read lighting symbols, and the accuracy may decrease in various drawing formats and environments with a lot of noise. In particular, it is difficult to perform learning and system updates to handle new lighting devices and symbols. When the shapes and line thicknesses, rotation angles, and deformation rates of the shapes stored in the database are different, it becomes difficult to recognize the lighting symbols in the drawing.

[0005] The present invention and its embodiments are proposed to solve the problems of the prior art described above. The object of the present invention and its embodiments is to provide a drawing analysis system that is unprecedentedly flexible and has excellent recognition performance, used for detecting equipment symbols, particularly lighting fixture symbols, contained in drawings, by appropriately combining deep learning with the application of equipment symbols in drawings, an automatic mapping function for LED lighting, and flexible data input and output functions. [Means for solving the problem]

[0006] The drawing analysis system of the present invention has the following configuration. (1) A drawing analysis system for detecting and classifying equipment symbols contained in drawings. (2) A dataset generation unit that generates multiple types of graphic datasets, including equipment symbols. (3) A learning unit that learns the features of equipment symbols based on the dataset using a deep learning model. (4) A detection unit that applies the deep learning model to an unknown dataset to automatically detect and classify multiple types of equipment symbols.

[0007] The drawing analysis system of the present invention preferably has the following configuration. (1) It is equipped with a transfer learning unit that uses a general object detection model as a pre-trained model and enables feature extraction specific to equipment symbols in drawings through transfer learning. (2) The system includes a confidence score assignment unit that assigns a confidence score to the detection results output by the deep learning model, and an accuracy improvement unit that improves the classification accuracy of equipment symbols based on the confidence score. (3) The system includes an identification unit that identifies existing lighting fixtures based on the detection results of equipment symbols, and a mapping unit that automatically searches a predefined database for the type of new LED lighting fixture corresponding to the detection results and maps each lighting fixture to an LED lighting fixture. (4) The system includes an interface that accepts image files in JPEG, PNG, and PDF formats as drawing data, a patching unit that converts the accepted drawing data into a standardized input format according to its size and resolution, and a detection unit that detects and classifies equipment symbols based on the standardized drawing data. (5) The system includes a cost calculation unit that calculates the cost of replacing lighting fixtures and the amount of energy saved based on the drawing analysis results, and a database unit that references data on the installation cost, power cost, and replacement cost of LED lighting fixtures corresponding to the detected equipment symbols. (6) The system includes an output unit that outputs the detected equipment symbols in visual and structured data formats, an overlay unit that highlights the location of the detected equipment symbols on the drawing, and a data output unit that outputs the detection results in JSON or CSV format to enable data integration with other systems. [Effects of the Invention]

[0008] According to the present invention and its embodiments, a drawing analysis system is provided that enables processing such as object detection using deep learning, transfer learning, mapping, cost calculation, and data output. [Brief explanation of the drawing]

[0009] [Figure 1] A functional block diagram illustrating an embodiment of the drawing analysis system of the present invention. [Modes for carrying out the invention]

[0010] [1. First Embodiment] [1-1. Configuration of the First Embodiment] Hereinafter, embodiments of the present invention will be specifically described with reference to the functional block diagram in Figure 1. The drawing analysis system of this embodiment applies deep learning technology using FasterR-CNN (Region-based Convolutional Neural Network) to automatically detect "lighting fixture symbols," which are one of the equipment symbols included in architectural drawings, with high accuracy. Traditionally, pattern matching and rule-based methods have been common for symbol recognition in architectural drawings. However, by utilizing FasterR-CNN, it becomes possible to automatically recognize a wide variety of symbols without requiring manual feature setting. Therefore, this embodiment has the following configuration.

[0011] (1) Dataset generation unit 1 The dataset generation unit 1 prepares training data for accurately recognizing equipment symbols placed within architectural drawings. In particular, to generate a dataset that can handle a wide variety of lighting fixture symbols (e.g., 100W and 20W straight fluorescent tubes, circular tubes, 6W and 40W incandescent bulbs, stairwell lights), the following procedure is followed.

[0012] (1-1) Collection of lighting fixture symbol images We will collect images containing lighting fixture symbols found in architectural drawings and design plans. This will involve obtaining samples from actual elementary school and similar facility drawing data, as well as from equipment symbol collections commonly used by designers. For each lighting fixture (e.g., 100W straight fluorescent tube, circular tube, etc.), we will collect a variety of samples to ensure variations in shape, angle, and size.

[0013] (1-2) Labeling Each symbol in the collected dataset is labeled based on its location and type using manual or semi-automatic tools. Specifically, each symbol is defined as representing a fixture, and then labeled according to category, such as "100W straight fluorescent lamp" or "circular fluorescent lamp."

[0014] (1-3) Data augmentation process The acquired data is augmented through processes such as rotation, flipping, color adjustment, and scaling to accommodate variations in symbol appearance and drawing quality. This allows the model to learn general-purpose features that enable it to adapt to various drawing environments.

[0015] (1-4) Structuring the dataset Structurally organize the labeled symbol data to match the input format of Faster R-CNN and construct a dataset that can be efficiently used during learning.

[0016] (2) Learning unit 2 In learning unit 2, the dataset constructed by dataset generation unit 1 is used to train the Faster R-CNN model to learn the features of symbols. In this process, the model acquires the feature quantities of equipment symbols and the information necessary for their identification, enabling it to handle unknown architectural drawings.

[0017] (2-1) Initial setting of the model and start of training Input the training data obtained by dataset generation unit 1 into the Faster R-CNN model. When using transfer learning, based on the weights pre-trained in a general object detection task, adjust the parameters according to the features of the equipment symbols in particular.

[0018] (2-2) Learning of feature quantities Extract the shape and feature quantities of each symbol in the convolutional layer (CNN) of the Faster R-CNN model and train it to be able to identify different features for each symbol. For example, to identify the differences in the shapes of straight tube type and circular fluorescent lamps, and the differences in the sizes of incandescent bulbs, optimize the model considering each feature.

[0019] (2-3) Loss function and backpropagation Calculate the loss function based on the recognition results of each symbol and adjust the parameters of the model so that the error is minimized. This improves the recognition accuracy. Through backpropagation, perform error backpropagation to update the weights of the model, and repeatedly train to enhance the recognition accuracy.

[0020] (2-**4**) Verification and accuracy improvement Using a validation dataset, the model's recognition accuracy is checked, and recognition performance is optimized while preventing overfitting and loss of accuracy. If there are many misrecognitions of specific lighting fixture symbols, retraining is performed to improve the model's accuracy, such as by generating additional data on the target symbols to enhance learning.

[0021] (3) Detection unit 3 The detection unit 3 uses the FasterR-CNN model built in the learning unit 2 to detect equipment symbols included in unknown architectural drawings and classify their types. This result is used for subsequent cost calculations and automatic mapping of LED lighting fixtures. It also communicates with the patching unit 8 (described later) by receiving its output.

[0022] (3-1) Preprocessing of unknown data To enable the FasterR-CNN model to accurately detect equipment symbols, newly input architectural drawing data (in JPEG or PNG format, etc.) is preprocessed so that the model can properly analyze it. Specifically, this involves processing such as scaling, denoising, and contrast adjustment.

[0023] (1) Scaling When drawing data has varying resolutions, unifying them to a consistent scale allows FasterR-CNN to accurately recognize the size of objects. By matching the input data resolution to the standard resolution used during training (e.g., 300 DPI), the model adjusts to match the scale it learned. This unified scale allows for consistent identification of equipment symbols of different sizes.

[0024] (2) Noise reduction This process removes noise such as small smudges and extraneous lines that occurred during scanning. Noise reduction filters, such as Gaussian filters, are applied to reduce unwanted fine dots and lines. Edge-preserving filters (such as bilateral filters) are applied to reduce noise while preserving the symbol's shape. Noise reduction makes the symbols sharper and improves detection accuracy.

[0025] (3) Contrast adjustment Architectural drawings may have insufficient contrast due to scanning or printing quality. Adjusting the contrast will make equipment symbols clearer on the drawing. Apply histogram equalization to uniformize the brightness of the entire drawing. Enhance the contrast as needed to make the outlines of symbols clearer, making it easier for the model to detect features.

[0026] (4) Binary conversion or grayscale conversion If architectural drawings contain unnecessary color elements, convert them to grayscale or binary. This highlights only the shape information essential for identifying symbols. Grayscale simplifies the image, allowing focus on symbolic shapes. Binary conversion enhances the contrast between the background and the object by setting a threshold to highlight symbols.

[0027] (3-2) Symbol position detection The FasterR-CNN model is applied to preprocessed architectural drawing data to detect the location of equipment symbols on the drawings. The model identifies the areas where equipment symbols exist within the architectural drawings and outputs their location information (bounding boxes).

[0028] (1) Feature extraction using FasterR-CNN The pre-processed image is input into the FasterR-CNN model, and features of equipment symbols within the image are extracted through convolutional layers (CNN). Feature extraction involves using convolutional filters to detect the shape and edge information of the symbols, which are then passed to the next layer.

[0029] (2) Generation of symbol candidate regions (RegionProposal) The RegionProposalNetwork (RPN) is used to generate areas (candidate areas) where equipment symbols may exist. Each candidate area is enclosed in a bounding box, has location information and size, and is assigned a score that predicts the presence or absence of a symbol. Many candidate areas are generated at this stage, but areas with low scores are narrowed down, and the analysis focuses on areas with high scores.

[0030] (3) Identification of equipment symbols The system classifies each candidate region and determines whether it is an equipment symbol. The classification layer of FasterR-CNN determines whether each candidate region belongs to a specific lighting fixture category, such as "100W straight fluorescent lamp" or "40W incandescent light bulb."

[0031] (4) Improving the accuracy of the bounding box Non-Maximum Suppression (NMS) is applied to merge overlapping bounding boxes into a single region. This reduces the detection of duplicate symbols and provides single positional information for each symbol. Finally, the bounding boxes of each symbol are placed on the drawing, and their positional coordinates and type are output.

[0032] (5) Output results Finally, the system obtains the location information (bounding box coordinates) and classification result (type of lighting fixture) for each detected equipment symbol. The output location information and classification results are used in the next steps, such as cost calculation and automatic mapping with LED lighting fixtures.

[0033] (3-3) Classification of Symbols (1) Setting classification criteria After equipment symbols are detected, FasterR-CNN outputs candidate symbol regions. Next, these candidate regions are classified to identify the type of lighting fixture. For example, lighting fixture types may include "100W straight fluorescent lamp" or "40W incandescent light bulb," and the features trained on the model are used for these classifications. The classification model learns features such as the shape, color, size, and placement of symbols, and determines which lighting fixture each symbol corresponds to. (2) Definition of lighting fixture categories To accurately classify each equipment symbol, lighting fixture categories (e.g., "100W straight fluorescent lamp," "circular fluorescent lamp," etc.) are defined in advance. Based on these categories, detected symbols are classified. In this process, the classification layer of FasterR-CNN plays a key role in determining which lighting fixture each candidate region belongs to. This classification enables efficient subsequent processing (e.g., selection of LED lighting fixtures and cost calculation). (3) Improvement of classification accuracy By utilizing deep learning, classification can be performed with higher accuracy than conventional methods. Conventional methods require manual definition of features and rules, which increases the risk of errors. However, by using FasterR-CNN, features can be automatically extracted from image data, enabling high-accuracy identification of equipment symbols.

[0034] (3-4) Calculation of confidence score (1) Method for calculating the confidence score For each detected symbol, FasterR-CNN calculates a confidence score. This score represents the probability that the model correctly classified that symbol. The score is usually expressed between 0 and 1, with a higher score indicating greater confidence. For example, if a 100W fluorescent tube is recognized with a probability of 90% or higher, its confidence score will be high. Conversely, a low confidence score suggests that the symbol may be misrecognized. (2) Filtering Symbols with high confidence scores are retained in the final classification results, while those with low confidence scores are excluded. For example, by including only symbols with a confidence score of 90% or higher in the final results, misclassification can be prevented. This improves the accuracy of the final symbol classification. Filtering can eliminate inaccurate recognitions, improving the reliability of the data. (3) Response to symbols with low confidence For symbols with low confidence levels (e.g., scores below 50%), a reclassification process or manual verification is implemented. This minimizes the risk of misrecognition. Reclassification can be improved by using additional training data.

[0035] (3-5) Output of classification results (1) Output format The location information and type of the finally classified symbols are output in a structured data format. Generally, JSON or CSV formats are used to organize the data for easier processing in other systems. For example, in JSON format, the type and location information (bounding box coordinates) of lighting fixtures can be organized as follows: { "symbol":"100W straight fluorescent lamp", "coordinates":[100,200,150,250] } This allows for efficient management of the location and type of equipment symbols, and facilitates data integration with other systems. (2) Highlighting of equipment symbols Highlighting the output results on the drawing makes it easier for designers and engineers to visually review the drawing. For example, a red border can be added to the drawing to visually highlight detected equipment symbols. This allows designers to see at a glance where the equipment symbols are located, making it easier to modify or adjust the design. (3) Convenience of data integration By outputting structured data, the effort required to integrate detection results into other systems is reduced. For example, when used for automatic mapping with LED lighting fixtures, the output data can be immediately imported into other software, allowing for smooth replacement of lighting fixtures and cost calculations. Furthermore, the data can be efficiently used in energy cost calculations and energy efficiency evaluations.

[0036] (4) Transfer Learning Section 4 The transfer learning unit 4 is a component that performs specialized learning of the features of equipment symbols based on a general object detection model. For example, in order to efficiently learn symbols with different shapes and sizes, such as 100W straight fluorescent lamps and 6W incandescent light bulbs, it performs the following processing. The transfer learning unit 4 is provided as part of the learning unit 2. Specifically, when the learning unit 2 learns equipment symbols using a deep learning model such as FasterR-CNN, the transfer learning unit 4 is the mechanism that utilizes the pre-trained weights of the general object detection model to perform specialized learning for architectural drawings.

[0037] (4-1) Outline of Transfer Learning Unit 4 The transfer learning unit 4 is provided as a functional addition to the learning unit 2, enabling applications from general object detection to architectural drawing analysis through transfer learning. The outline of its processing is as follows. (1) Model initialization The pre-trained weights of a general object detection model (e.g., FasterR-CNN) are loaded, and parameters suitable for general object detection are set as initial values. (2) Adjustment specifically for architectural drawings The learning parameters are updated in the intermediate layers of the model to emphasize specific shapes and edges, taking into account the characteristics of 100W straight fluorescent tubes and circular tubes. (3) Iterative learning process In cases where visual features overlap, such as stairwell lights or 40W incandescent light bulbs, the error is adjusted to minimize the loss function, enabling optimal classification through multiple training iterations.

[0038] (4-2) Configuration of the transfer learning unit 4 The transfer learning unit 4 consists of the following elements. (1) Module for obtaining pre-trained models We will utilize deep learning models such as FasterR-CNN and YOLO, which were trained for general object detection, as a base. For example, we will start customizing them for architectural drawing analysis by leveraging weights trained on the COCO dataset. (2) Learning module specializing in architectural drawings To adapt to the characteristics of architectural drawings, the intermediate and output layers are set to a state where they can be retrained. Data based on new equipment symbols (e.g., 100W straight fluorescent lamp, 6W incandescent bulb) is input, and the existing model is fine-tuned for specific tasks. (3) Retrainable intermediate layer The convolutional layers and fully connected layers near the end of the pre-trained model are made retrainable to learn the characteristics of a new task. For example, the "circular" and "rectangular" features detected by the existing model can be applied to the recognition of "circular tubes" and "straight fluorescent lamps." (4) Loss calculation module Errors are calculated using cross-entropy loss (for classification tasks) or localization loss (for detection tasks). Backpropagation is used to minimize errors, and training is repeated. (5) Hyperparameter adjustment module By optimizing parameters such as the learning rate and mini-batch size, efficient and highly accurate learning is achieved.

[0039] (4-3) Processing of the transfer learning unit 4 The processing of the transfer learning unit 4, which has the configuration described above, is carried out in the following steps. (1) Loading a pre-trained model • For widely used large datasets such as the COCO dataset, load a pre-trained model (e.g., FasterR-CNN). The model has the ability to extract features from common objects (dogs, cats, cars, etc.). (2) Settings for relearning • Fixing and releasing layers The initial layers (e.g., layers involved in edge detection) are fixed and excluded from retraining. This preserves the basic feature extraction capabilities. The final layers (e.g., classification layers, bounding box adjustment layers) are set to be retrainable and adapted to the characteristics of architectural drawings. (3) Input of training data • Input architectural drawing data including equipment symbols (e.g., 100W straight fluorescent tube, circular tube). The dataset must be labeled (e.g., "100W straight fluorescent lamp", "staircase light"). (4) Feature retraining • Processing in the convolutional layer Image data of architectural drawings is input into a convolutional layer, and features are extracted. For example, the following features are learned. "Long straight line" = straight fluorescent lamp "A series of curves" = circular tube "Small, dot-like shape" = stairwell lamp • Optimization in the intermediate layer The retrainable intermediate layer optimizes features that distinguish subtle shapes and edges within the drawing. For example, it specializes in distinguishing between "straight fluorescent tubes" and "incandescent light bulbs." (5) Calculation of the loss function and backpropagation • Classification task If a "100W straight fluorescent tube" is misclassified as a "circular tube," the loss is calculated and the model parameters are adjusted. • Detection task If the bounding box (position information) is inaccurate, the error is calculated to improve the accuracy of position adjustment. (6) Evaluation of learning results • Use a validation dataset to evaluate whether the model can accurately detect and classify symbols. • If the accuracy is low, the model will be retrained. (7) Saving the trained model • Save the trained model and make it available for use in the architectural drawing analysis process.

[0040] (4-4)Specific examples The following is an example of a specific learning process performed in the transfer learning unit 4. (1) Input image Architectural drawing data including 100W straight fluorescent tubes, circular fluorescent tubes, and stairwell lights. (2) Output of the initial model A "straight fluorescent tube" is recognized as a "linear object." The "circular tube" is recognized as a "circular object". The stairwell light is being misidentified. (3) Output after transfer learning "Straight fluorescent tubes" are identified by their length-to-width ratio and their angle of inclination. "Circular tubes" are classified based on their circular shape and the arrangement pattern inside. "Staircase lights" are accurately recognized based on the combination of their shape and size. (4) Points By incorporating a transfer learning unit 4, the system efficiently learns features specific to architectural drawings while utilizing existing pre-trained models. It possesses versatility that allows it to adapt to variations in different architectural drawings and equipment symbols.

[0041] (5) Accuracy improvement part 5 The accuracy improvement unit 5 is provided as part of the detection unit 3. The accuracy improvement unit 5 is responsible for calculating a confidence score of the output results during the process in which the detection unit 3 analyzes unknown architectural drawing data and detects and classifies equipment symbols, and for improving accuracy based on that score. The accuracy improvement unit 5 is provided as an auxiliary function aimed at improving the accuracy of the classification results in the detection unit 3. The accuracy improvement unit 5 has a function to improve the classification results based on a confidence score. For example, in order to accurately distinguish between 100W and 20W straight fluorescent lamps, the following processing is performed.

[0042] (1) Assignment of confidence score For each detected symbol, the model calculates a confidence score based on the probability that it recognized it as belonging to a specific category (e.g., a 6W incandescent light bulb, a circular tube). A higher score indicates a lower risk of misrecognition. (2) Filtering by score For example, misclassification can be reduced by including only symbols with a confidence level of 90% or higher in the classification results and excluding symbols with a confidence level of less than 50%. (3) Implementation of relearning If many low-scoring 40W incandescent light bulb symbols appear, the system generates additional data related to those symbols and retrains the model to improve accuracy.

[0043] (5-1) Configuration of the precision improvement unit 5 The accuracy improvement unit 5 is responsible for evaluating the reliability of the detected equipment symbols and improving the final classification result, and its configuration is as follows. (1) Confidence score calculation module This module calculates a confidence score (e.g., a probability value between 0 and 1) for each symbol recognition result output from object detection models such as FasterR-CNN and YOLO. For example, when a "100W straight fluorescent lamp" is detected, the reliability of the recognition result is scored (e.g., 0.95). (2) Score-based filtering module A module that excludes low-scoring results based on confidence scores. Set a score threshold (e.g., 0.8 or higher) to eliminate symbols with a high risk of misclassification. (3) Reclassification module A module that attempts to reclassify items with low scores. For example, if the shapes are similar, such as "straight fluorescent tubes" and "circular fluorescent tubes," they will be reclassified using additional features. (4) Retraining support module We collect misclassified results and low-scoring data, and construct a dataset to be used for subsequent retraining. (5) Output accuracy optimization module A module that checks the accuracy after scoring and optimizes the output of the final result. The output will only include data with high scores.

[0044] (5-2) Processing of the precision improvement unit 5 (1) Calculation of confidence score • Input data: The detection results of equipment symbols output from detection unit 3 (FasterR-CNN, etc.). For example, when "100W straight fluorescent lamp" is detected in the architectural drawing, its recognition confidence is evaluated as 0.93. • Confidence score calculation: The model obtains a confidence score based on the probability value it uses to classify the detection results. For example, if the score for "circular pipe" is 0.87 and the score for "staircase light" is 0.72, the model will determine that the circular pipe is more reliable. (2) Filtering based on score criteria • Setting thresholds: Only results with a high confidence score (e.g., 0.8 or higher) are included in the final result. "Staircase lights" with a score of less than 0.8 are excluded due to the high possibility of misrecognition. Exception handling: Even if the score is low, we will attempt to re-evaluate it based on specific shapes or positional patterns (e.g., symbols placed in a specific area of ​​the drawing). (3) Reclassification process • Re-evaluation of shape characteristics Symbols that were excluded due to low scores (e.g., "straight fluorescent tube" with a score of 0.6) will be re-evaluated for their shape and size characteristics. For example, the length-to-width ratio and inclination characteristics of "straight fluorescent tube" will be recalculated and reclassified. Considering the context: To prevent misidentification, consider the relative placement and density of surrounding symbols. For example, if stairwell lights are frequently placed in a particular area, prioritize their consideration. (4) Construction of a dataset for retraining • Collection of misclassification data We collected false positives and low-scoring results to build a dataset for retraining. For example, we added cases where a "40W incandescent light bulb" was mistakenly identified as a "6W incandescent light bulb" to the dataset. • Additional data expansion: For symbols that are frequently misrecognized (e.g., circular tubes), the data is augmented by rotation, inversion, and noise addition, and this is reflected in the learning process. (5) Optimization of output accuracy ·Deduplication: If multiple detection results overlap (e.g., the same symbol is detected multiple times), the one with the higher confidence score is retained. For example, Non-MaximumSuppression (NMS) can be applied to merge bounding boxes. • Organizing the final output results: Output as structured data (e.g., JSON format) containing only high-scoring symbols.

[0045] (5-3)Specific examples (1) Scenario • Input drawing: Circular tubes (score 0.85), straight fluorescent tubes (score 0.92), and stairwell lights (score 0.65) were detected. • Processing details: The stairwell light (score of 0.65) was excluded for the time being because it fell below the threshold (0.8 or less). The characteristics of the stairwell lights were re-evaluated (in conjunction with the surrounding placement patterns), and the score was raised to 0.78, which was then provisionally adopted. The straight fluorescent tubes and circular tubes were retained due to their high scores. • Output result: "Circle pipe: Score 0.85" "Fluorescent tube: Score 0.92" "Staircase light: Score 0.78"

[0046] (5-4) Effects of the precision improvement unit 5 • Eliminate misrecognition: Remove low-confidence symbols to improve the accuracy of the final result. • Flexibility in re-evaluation: Reclassification and additional learning improve the accuracy of identifying specific, difficult symbols. • Improved final output quality: By transferring high-precision data to other systems, the burden on designers and engineers is reduced.

[0047] (6) Mapping section 6 The mapping unit 6 is provided as a subsequent processing step after the detection unit 3. It is provided with the added function of mapping existing lighting fixtures to new LED lighting fixtures based on the detection results of equipment symbols obtained by the detection unit 3. The mapping unit 6 is added to utilize the results output by the detection unit 3 in the next process (e.g., equipment selection and replacement planning), and has the function of mapping existing lighting fixtures to LED lighting fixtures. For example, when mapping stairwell lights to energy-saving LED fixtures, the following processing is performed. • Identification process The identification unit determines whether the detected symbol is a stairwell light or a 100W straight fluorescent lamp. • Search for compatible lighting fixtures Based on the identification results, mapping information such as a 100W straight fluorescent lamp corresponding to a 50W straight LED tube is generated. Search from Tabes. • Registering results The system lists each symbol and its corresponding LED fixture information, outputting it in a format that can be used for subsequent cost calculations.

[0048] (6-1) Configuration of the mapping unit 6 The mapping unit 6 consists of components that associate detected equipment symbols with existing lighting fixtures and further map them to new LED lighting fixtures based on a predefined database unit 9. (1) Lighting fixture identification module Based on the classification results of the equipment symbols output from the detection unit 3, the module identifies existing lighting fixtures (e.g., 100W straight fluorescent lamps, circular tubes) accurately by utilizing shape, size, and arrangement information. (2) Lighting fixture mapping module This module searches for LED lighting fixtures corresponding to identified existing lighting fixtures based on the database unit 9 and performs appropriate mapping. For example, it retrieves information from the database unit 9 that a 100W straight fluorescent lamp corresponds to a 20W straight LED lamp. (3) Database Unit 9 Connection Module This module is responsible for connecting to the database unit 9, which stores the specifications and performance (e.g., power consumption, luminous efficiency) of LED lighting fixtures. It efficiently executes queries to the database unit 9 and retrieves the results. (4) Output result generation module A module that visually displays mapping results on architectural drawings, or outputs them as structured data in JSON or CSV format.

[0049] (6-2) Processing of Mapping Unit 6 The processing of the mapping unit 6 consists of the following steps. (1) Identification of equipment symbols • Input data The classification result of the symbols obtained by the detection unit 3 (e.g., "100W straight fluorescent lamp"). • Identification process Based on the shape and size characteristics of the symbols, the system identifies the corresponding existing lighting fixtures. For example, a detected symbol with a "linear shape and constant length" is identified as a "linear fluorescent lamp." In the case of complex symbols, it can also identify small shapes such as stairwell lights and circular symbols (circular tubes). (2) Mapping of lighting fixtures • Execution of Query 9 in the Database Section: The lighting fixture identification module searches for the type of LED lighting fixture that corresponds to the existing lighting fixture identified. For example, it can map a 100W straight fluorescent lamp to a 20W straight LED tube, and a circular tube to an LED circular tube. • Obtaining the results of the correspondence: Database section 9 retrieves information such as the model number, power consumption, and installation cost of LED lighting fixtures. For example, for a "20W straight-tube LED," the corresponding information would include a cost of 3000 yen and a luminous efficiency of 120 lm / W. (3) Integration of mapping results • Integration of mapping information: The detected equipment symbols and corresponding LED lighting fixture information are integrated to establish one-to-one or one-to-many correspondences. For example, as follows: Detection: 100W straight fluorescent lamp Mapping: 20W straight LED tube (cost 3000 yen, power consumption 20W) ·Deduplication: If the same lighting fixture is detected multiple times, the location information is integrated to reduce duplicate results. (4) Generation of output results • Visual output The mapping results are visualized on the architectural drawings. For example, a "straight fluorescent tube" is enclosed in a red frame on the drawing, and the annotation "LED straight tube 20W" is added next to it. • Output of data format Structured data can be output in JSON or CSV format and prepared in a way that allows it to be integrated with other systems (e.g., cost calculation software or energy efficiency evaluation systems).

[0050] (6-3)Specific examples (1) Scenario • Input architectural drawings "100W straight fluorescent tubes" and "circular fluorescent tubes" were detected. (2) Processing details 1. Identifying lighting fixtures Detection result: "100W straight fluorescent lamp" (shape: rectangular, length: 120cm). Detection result: "Circular tube" (Shape: circular, Diameter: 30 cm). 2. Database Section 9 Search "100W straight fluorescent lamp" → 20W straight LED tube type (Cost: 3000 yen, Power consumption: 20W, Lifespan: 50,000 hours). "Circle tube" → LED circle tube (Cost: 4000 yen, Power consumption: 18W, Lifespan: 45000 hours). 3. Integration of results: The detected symbols and LED lighting fixture information are integrated to create a one-to-one correspondence. 4. Output generation: Each symbol is highlighted on the drawing, and corresponding LED lighting fixture information is displayed as an annotation. Output the results in JSON format.

[0051] (6-4) Effects of the mapping unit 6 • Selection of efficient lighting fixtures: By automatically suggesting the optimal LED lighting fixture based on the detection results, the burden on designers is reduced. • Maximizing energy-saving effects: Based on existing power consumption data, the effectiveness of switching to LED lighting fixtures can be quickly evaluated. • Convenience of data integration: Providing detection results and mapping information as structured data makes it easy to use in other systems.

[0052] (7) Calculation section 7 The calculation unit 7 is responsible for calculating replacement costs, energy savings, and payback periods based on the results of architectural drawing analysis and replacement mapping. The results from the calculation unit 7 are used by the output unit as reports and annotations on drawings. (7-1) Configuration of the calculation unit 7 The calculation unit 7 consists of modules that calculate replacement costs, energy savings, and the payback period based on the mapping results of replacement parts and information from the detection unit 3. This module consists of the following submodules. (1) Replacement cost calculation module: Based on the number of detected equipment, it calculates the total of the unit price of replacement parts and installation costs. (2) Energy Reduction Calculation Module: Calculates the annual energy reduction and cost savings based on the difference in power consumption between the current equipment and the replacement equipment, and the annual usage time. (3) Payback period calculation module: The payback period is calculated by dividing the replacement cost by the reduction cost. (7-2) Processing of the calculation unit 7 The calculation unit 7 performs the following processing based on the information received from the detection unit 3 and the mapping unit 6. (1) Obtain the number of detected lighting fixtures and the mapping results. (2) For each symbol, the replacement cost, annual reduction, and payback period are calculated sequentially. (3) Output the results in a format that can be used in the next step. (7-3) Specific Examples (1) When replacing 50 "100W fluorescent lamps" with "20W LED straight tubes" (a) Replacement cost: ¥3,000 (LED price) + ¥1,000 (installation cost) × 50 locations = ¥200,000. (b) Reduction amount: power consumption difference 80W x 50 units x 3,000 hours = 240,000Wh (240kWh). (c) Cost reduction: 240kWh × 0.10 / kWh = 120,000. (d) Payback period: ¥200,000 ÷ ¥120,000 = approximately 1.7 years.

[0053] (8) Patching section 8 The patching unit 8 is provided as a pre-processing step for the dataset generation unit 1. It plays a role in assisting dataset generation, learning, and detection by converting the input architectural drawing data into a standardized format. In this process, the detection unit 3 cooperates by receiving the output of the patching unit 8. The patching unit 8 and the detection unit 3 are responsible for standardizing the input data and detecting equipment symbols. For example, to accurately detect a 100W straight fluorescent lamp in an architectural drawing in PDF format, the following processing is performed. (1) Standardization of image data If the architectural drawings are in JPEG format, they will be converted to the standard resolution of the trained model (e.g., 300 DPI). Additionally, the contrast of the drawings will be adjusted to ensure that details such as stairwell lights and circular pipes are clearly visible. (2) Model application Standardized data is input into the FasterR-CNN model, and the bounding box coordinates of each equipment symbol are output. (3) Classification of results For example, the classification layer of the model determines whether the symbol within the bounding box is a 100W straight fluorescent lamp or a 40W incandescent light bulb.

[0054] (8-1) Configuration of patching section 8 The patching unit 8 standardizes the format and size of the architectural drawings and performs preprocessing to convert them into a state that can be analyzed by the deep learning model (detection unit 3). (1) Format conversion module This module converts architectural drawings provided in different formats, such as PDF, JPEG, and PNG, into a format that the model can analyze (e.g., a standard resolution image format). (2) Size and resolution adjustment module This module converts architectural drawings to a unified scale (e.g., 300 DPI, fixed size) if their size or resolution differs from that used during model training. It performs upscaling for lower resolutions and downsampling for higher resolutions. (3) Noise reduction module This module removes smudges and unnecessary lines during scanning. It applies Gaussian filters and edge-preserving filters to clarify the symbolic areas. (4) Contrast adjustment module This module adjusts the contrast of faint lines or text to make the image easier for the model to analyze. It also performs histogram equalization and dynamic range adjustment. (5) Grayscale-to-binary conversion module A module that converts color architectural drawings to grayscale or binary format to remove information unnecessary for symbol recognition.

[0055] (8-2) Processing of patching section 8 (1) Format conversion The system uses architectural drawings in PDF, JPEG, and PNG formats as input data. PDFs are converted to images page by page. JPEG and PNG files are imported directly. (2) Unification of size and resolution Low-resolution images (e.g., 150 DPI) are upscaled to 300 DPI. If the size is extremely large, it is resized to a standard analysis range (e.g., 1024 x 1024 pixels). (3) Noise reduction A Gaussian filter removes imperfections (dots and line segments) generated during scanning. This suppresses noise while preserving the edges of the symbols. (4) Contrast adjustment The brightness of the entire drawing is made uniform, and thin lines and symbols are made to stand out. In particular, the shapes of "straight fluorescent tubes" and "circular tubes" are adjusted so that they can be clearly recognized. (5) Grayscale to Binary Conversion Grayscale conversion enhances shape information. Binary conversion removes background noise and highlights only the symbols.

[0056] (8-3) Coordinated processing of patching unit 8 and detection unit 3 (1) Input of preprocessing results The detection unit 3 receives the image data standardized by the patching unit 8. (2) Feature extraction Through convolutional layers, features (straight lines, circles, size, arrangement) within architectural drawings are extracted. For example, "100W straight fluorescent tube" is recognized as a "long straight shape." (3) Generation of symbol candidate regions The RegionProposalNetwork (RPN) identifies areas where equipment symbols are likely to exist and generates bounding boxes (area information). (4) Symbol Classification Each candidate area is passed through a classification model to determine categories such as "100W straight fluorescent lamp," "staircase light," and "circular tube." (5) Bounding box adjustment Duplicate bounding boxes are merged to improve the accuracy of location information. (6) Output of detection results The detection results are output as structured data (e.g., JSON format). For example, as shown below. json { "symbol":"100W straight fluorescent lamp", "coordinates":[100,200,150,250] }

[0057] (8-4)Specific examples The following is a specific example illustrating the coordinated processing between the patching unit 8 and the detection unit 3. (1) Input data Architectural drawings in PDF format (300 DPI, including 100W fluorescent tubes and stairwell lighting). (2) Processing flow Patching section 8 converts the PDF to an image format, adjusts the resolution to 300 DPI, and sharpens the shape of the "straight fluorescent lamp" through noise reduction and contrast adjustment. Unnecessary background information is removed by converting to grayscale. The detection unit 3 extracts the characteristics of the straight fluorescent lamp through the convolutional layer and generates a bounding box using the RegionProposalNetwork (e.g., location information [100,200,150,250]). Then, the classification layer determines that it is a "100W straight fluorescent lamp". The detection result is output, for example, in JSON format and highlighted on the drawing.

[0058] (8-5) The effect of the cooperation between the patching unit 8 and the detection unit 3 The patching unit 8 unifies the diverse formats and resolutions of architectural drawings, improving analysis accuracy. It also prevents misrecognition by the detection unit 3 by eliminating noise. The detection unit 3 utilizes a high-precision deep learning model to accurately detect and classify equipment symbols, integrates the bounding box and classification results, and provides a foundation for drawing analysis.

[0059] (9) Database Section 9 The database unit 9 is responsible for providing information to calculate replacement costs and energy savings based on the detected equipment symbols. It also stores information on lighting fixture specifications and corresponding LED lighting fixtures, and provides it in a format usable in subsequent processes. Therefore, the database unit 9 manages the information used for cost calculation and energy efficiency evaluation. For example, when calculating the replacement cost of a 20W straight fluorescent lamp, the following process is performed. (1) Reference of cost information Refer to database section 9, which stores the installation cost, replacement cost, and power consumption of each lighting fixture (e.g., 6W incandescent light bulb, circular tube). (2) Calculation of energy savings By replacing lighting fixtures with LEDs, the amount of savings is calculated based on the difference in power consumption, for example, when changing a 100W fluorescent lamp to a 20W LED. (3) Provision of results Output the calculation results in a format that is easily usable by other systems.

[0060] (9-1) Configuration of Database Section 9 The database section 9 consists of the following components: (1) Lighting fixture specification data module Database unit 9 records the power consumption, installation cost, lifespan, etc., of detected lighting fixtures (e.g., 100W straight fluorescent tubes, circular tubes). Examples of items to store: • Name of lighting fixture (e.g., 100W straight fluorescent lamp) • Power consumption (e.g., 100W) • Installation fee (e.g., ¥3000) ·Use life (e.g. 20,000 hours) (2) Data module compatible with LED lighting fixtures Stores data for mapping LED lighting fixtures to existing lighting fixtures. Examples of items to store: • Compatible LED fixture name (e.g., 20W straight tube LED) ·Power consumption (e.g. 20W) • Installation fee (e.g., ¥4000) • Luminous efficiency (e.g., 120 lm / W) (3) Cost calculation data module Data used to calculate replacement costs and energy savings based on detected equipment symbols and corresponding LED lighting fixtures. Examples of items to store: • Replacement cost (unit price of the equipment + installation cost) • Energy savings (calculated from the difference between annual usage time and power consumption) • Payback period (the period during which initial costs are recovered through savings) (4) Query execution and data acquisition module A module equipped with a search function for efficiently obtaining information on lighting fixtures and LED lighting fixtures. (5) Data update module A module that updates the information in database section 9 when new LED lighting fixtures or replacement costs are added.

[0061] (9-2) Processing of Database Section 9 (1) Data storage • Data storage for lighting fixtures: Information about existing lighting fixtures (e.g., power consumption and lifespan of a 100W straight fluorescent lamp) is registered in the database unit 9. • Data storage for LED lighting fixtures: Information on LED lighting fixtures compatible with existing fixtures is registered and used for mapping. (2) Search for lighting fixture information • Input data: Symbol information (e.g., 100W straight fluorescent lamp) passed from the detection unit 3 or the mapping unit 6. • Processing details: The system searches for lighting fixture data within database section 9 and retrieves the corresponding fixture information (e.g., 20W straight tube LED). (3) Calculation of replacement costs and energy savings • Calculation of replacement costs: Unit price (e.g., LED tube light ¥4000) + installation fee (e.g., ¥2000). Total replacement cost: ¥6000. • Calculation of energy savings: Power consumption difference: 100W (existing) - 20W (LED) = 80W. Annual savings: 80W x annual usage time (e.g. 3000 hours) = 240,000Wh (240kWh). • Calculation of the payback period: It is calculated by dividing the initial cost by the annual savings (electricity bill). Example: 6000 ÷ (27 / kWh × 240kWh) = approximately 0.93 years (approximately 11 months). (4) Output of data The calculation results are output in JSON or CSV format and can be linked with cost calculation systems and energy saving evaluation systems. Example output: json { "original_fixture":"100W straight fluorescent lamp", "led_fixture":"Straight tube LED20W", "exchange_cost":6000, "annual_saving":240, "payback_period":0.93 } (5) Data update When new LED products are introduced to the market, the information is added to the database section 9. The database is also updated as needed if replacement costs or installation unit prices change.

[0062] (9-3) Specific Examples The following is a specific example illustrating the processing of the database section 9. (1) Input data: Detection unit 3 detected "100W straight fluorescent lamp" and planned to replace it. (2) Processing flow • Obtain information on "100W straight fluorescent lamp" from the lighting fixture specifications data module. • Obtain information on the "20W straight tube LED" compatible with the LED lighting fixture data module. The cost calculation data module calculates the replacement cost (¥6000), annual savings (240kWh), and payback period (approximately 11 months). • Output the calculation results in JSON format, for example, as shown below. json { "original_fixture":"100W straight fluorescent lamp", "led_fixture":"Straight tube LED20W", "exchange_cost":6000, "annual_saving":240, "payback_period":0.93 }

[0063] (9-4) Effects of Database Section 9 (1) Prompt information provision: Information on lighting fixtures and LED lighting fixtures is centrally managed, allowing for smooth processing in the mapping unit 6 and detection unit 3. (2) Highly accurate cost calculation: We quantify the cost of replacing lighting fixtures and the energy-saving effects, providing designers and engineers with concrete plans. (3) Adaptability through data updates: Even when new lighting fixtures are introduced, the system can be adapted long-term by updating the data.

[0064] (10) Data output unit 10 The data output unit 10 provides a function to format detection results and analysis results into a format that users can utilize and visually confirm. It also handles conversion to data formats for linking with other systems. The data output unit 10 provides detection results in both structured and visual formats. For example, when highlighting the location of stairwell lights on an architectural drawing, it performs the following processing. (1) Visual display The locations of detected circular tubes and 40W incandescent light bulbs are highlighted on the drawing with red borders or icons. (2) Output of data format The detection results are organized in JSON format, recording, for example, the location coordinates and type of a "100W straight fluorescent lamp". (3) Data Integration Using structured data, we will integrate it with systems for planning and calculating the costs of LED lighting fixture installations.

[0065] (10-1) Configuration of the data output unit 10 The data output unit 10 is provided as a post-processing step for the detection unit 3, mapping unit 6, and database unit 9. Its role is to visually display the information obtained from these units or to output it in a format that can be integrated into other systems. The data output unit 10 consists of the following components. (1) Structured data generation module This module formats the results from the detection unit 3 and the database unit 9 into structured data formats such as JSON and CSV. Examples of items to store: • Symbol name (e.g., 100W straight fluorescent lamp) • Detection position (bounding box coordinates) • Mapping results (Example: 20W LED tube) (2) Visual output generation module A module that visually displays detection results on architectural drawings. Example: A symbol on the drawing is enclosed in a red frame, and the symbol name and information about the corresponding LED lighting fixture are displayed next to it as an annotation. (3) Overlay generation module It has a feature that adds an overlay to architectural drawings to make the detection results stand out. Example: Overlay a colored frame or icon on the location of the detected symbol. (4) Data format conversion module A module that converts output data into different formats (such as JSON and CSV) to prepare it for use by external systems. (5) Data Integration Module A module that converts and integrates data into a format suitable for use with other analysis systems and design tools (such as energy calculation software).

[0066] (10-2) Processing of the data output unit 10 (1) Obtaining detection results The system acquires symbol information (e.g., 100W straight fluorescent lamp, stairwell light) sent from the detection unit 3 as input data. Information items include, for example, the symbol name, detection location (coordinates), and mapping results. (2) Processing details: (a) Conversion of detection results Based on the input information, it is converted into a user-friendly format. For example, detection results can be converted to JSON or CSV format. ·JSON format example: json { "symbol":"100W straight fluorescent lamp", "coordinates":[100,200,150,250], "mapped_fixture":"LED straight tube 20W", "exchange_cost":6000 } ·CSV format example: Symbol name, X1, Y1, X2, Y2, Compatible LED fixture, Replacement cost Fluorescent tubes 100W, 100W, 200W, 150W, 250W; LED tube 20W; 6000 yen (b) Generation of visual output Colored frames and annotations are added to the detected locations on the architectural drawings. Example: The red frame encloses "100W straight fluorescent lamp," and the annotation "20W LED straight tube (replacement cost ¥6000)" is displayed to the right. (c) Overlay generation Symbols detected on the drawing image are highlighted. Example: Overlay a yellow icon on the location of the stairwell light. The position of each symbol is highlighted based on the bounding box coordinates. (d) Data format conversion Converts data to the user's requested format (e.g., JSON, CSV) and provides it as exportable data. The output format can be flexibly configured (e.g., a format that includes only specific data items). (e) Data integration Data integration with other systems (e.g., energy efficiency evaluation systems and design support software). Example: Export detection results in CSV format and import them directly into the energy efficiency evaluation tool.

[0067] (10-3) Specific Examples The following is a specific example illustrating the processing of the output section. (1) Input data: The detection unit 3 detects "100W straight fluorescent lamp" and "circular tube," and provides bounding box coordinates and mapping information. Detection locations: [100, 200, 150, 250] (straight fluorescent tubes), [300, 400, 350, 450] (circular tubes) • Mapping results: 20W straight LED tube (straight fluorescent lamp), 18W circular LED tube (circular tube) (2) Generating structured data: For example, generate data in JSON format. json [ { "symbol":"100W straight fluorescent lamp", "coordinates":[100,200,150,250], "mapped_fixture":"LED straight tube 20W", "exchange_cost":6000 }, { "symbol": "Circle tube", "coordinates":[300,400,350,450], "mapped_fixture":"LED Circle 18W", "exchange_cost":7000 } ] (3) Generation of visual output: The symbol's location is highlighted on the drawing. • Fluorescent tube: Red frame + note "LED straight tube 20W (¥6000)". • Circle tube: Blue frame + note "LED Circle 18W (¥7000)". (4) Data format conversion: Convert and save in CSV format: • Symbol name, X1, Y1, X2, Y2, Compatible LED fixture, Replacement cost • Fluorescent tubes 100W, 100W, 200W, 150W, 250W; LED tube 20W; 6000 yen • Circle tube, 300, 400, 350, 450, LED circle 18W, 7000 yen

[0068] (10-4) Effects of the data output unit 10 (1) Visual clarity: Detected symbols are highlighted on architectural drawings, making them easy for engineers and designers to identify. (2) Convenience with structured data: Outputting in JSON or CSV format allows for seamless integration with other systems. (3) Flexibility of data integration: Data preparation for import into other energy efficiency evaluation systems and design support tools is easy.

[0069] [1-2. Operation of the Embodiment] The following describes the equipment symbol detection and classification process using FasterR-CNN.

[0070] 1. Equipment Symbol Learning Process When using a FasterR-CNN model to detect and classify equipment symbols, the model is trained to correctly identify equipment symbols by providing a wide variety of equipment symbol shapes and names as training data. This training process proceeds as follows.

[0071] (1) Preparation of training data The system collects data on various equipment symbols included in architectural drawings (e.g., 100W straight fluorescent tube, 20W circular tube, 40W incandescent light bulb, etc.) and labels each symbol. (1) Data collection • Collect actual equipment symbols placed on drawings, and gather images that include the characteristic shapes and sizes of each symbol. • Provide a wide variety of equipment symbols to accommodate different drawing formats, scales, and angles. (2) Labeling • Label each symbol's image with the corresponding name, such as "100W straight fluorescent tube" or "20W circular tube." • For labeling, use a specialized labeling tool (e.g., LabelImg) to draw a bounding box surrounding the area containing the symbol and assign a name to it.

[0072] (2) Learning of features The FasterR-CNN model learns the features of equipment symbols based on the shape and labels of the training data. During this process, the specific shape, edges, and proportions of the symbols are accumulated as features of the model. (1) Feature extraction • The convolutional layer (CNN) extracts the shape and edge information of each symbol, and the learning process progresses so that the features of each symbol are understood. For example, a "100W straight fluorescent tube" with a distinctive linear shape and a "20W circular tube" with a hollow, circular shape will be trained in the model to learn different features, allowing the model to distinguish between them. (2) Learning the criteria for discrimination • For a symbol like a "40W incandescent light bulb," its round shape and internal structure are distinctive features, and these characteristics are stored as a basis for the model. The model accumulates these features along with the label information for each symbol, so even for unknown drawings, if a similar shape is found, the corresponding name can be predicted.

[0073] (3) Reasoning (Application to unknown architectural drawings) When a FasterR-CNN model is applied to new architectural drawings, the model classifies each symbol based on the learned features. (1) Classification of candidate regions • If the detected candidate region has features such as "straight line shape + rounded edges," the model predicts it as a "100W straight fluorescent tube." Similarly, for other shapes, the system determines which equipment symbol each candidate region corresponds to based on the learned features.

[0074] (4) Correspondence of new symbols If new symbols are added in the future, or if the model's accuracy needs improvement, it is possible to create additional training data and retrain the model. Retraining will allow the model to handle the new symbols and improve its recognition accuracy. In this way, the FasterR-CNN model determines equipment symbols based on the shape-name combinations given in the training data, and automatically identifies the type and location of symbols even in new architectural drawings.

[0075] 2. Process for detecting equipment symbols included in architectural drawings Let's assume that a building drawing contains numerous lighting fixtures such as "100W straight fluorescent tubes," "20W circular tubes," and "40W incandescent light bulbs." This drawing is scanned and imported into the system as image data (such as JPEG format), and the system automatically detects and classifies the lighting fixtures.

[0076] (1) Feature extraction using FasterR-CNN Pre-processed architectural drawing data is input into the FasterR-CNN model, and first, features of equipment symbols are extracted through a convolutional layer (CNN). (1) Assume that the drawing contains multiple symbols for both "100W straight fluorescent lamps" and "40W incandescent light bulbs". (2) In the convolutional layer, features such as the shape, edge pattern, line thickness, and angle of each equipment symbol are extracted. (3) For example, the symbol for a "100W straight fluorescent lamp" has a linear shape with rounded ends, while the symbol for a "40W incandescent light bulb" is often round with a darker center. (4) These features are then transferred to the next step (generation of symbol candidate regions).

[0077] (2) Generation of symbol candidate regions (RegionProposal) Once features are extracted, the RegionProposalNetwork (RPN) works to generate regions (candidate regions) that are highly likely to contain equipment symbols. (1) Within the drawing, a bounding box is created to enclose the area where it is predicted that "100W straight fluorescent lamps" and "40W incandescent light bulbs" will be included. (2) Each candidate region is enclosed as a "bounding box," and a probability (score) of the presence of an equipment symbol in that region is assigned. (3) For example, if a vertically elongated candidate region is generated for the symbol "100W straight fluorescent lamp" and the score is 90%, the model determines that there is a high probability that a "100W straight fluorescent lamp" exists in that region. (4) At this stage, regions with low core scores are excluded from the analysis, and candidate regions with high scores are processed preferentially.

[0078] (3) Identification of equipment symbols The system classifies each candidate region and determines the type of equipment symbol. The classification layer of FasterR-CNN categorizes each candidate region into a specific lighting fixture category, such as "100W straight fluorescent tube," "20W circular tube," or "40W incandescent light bulb." (1) For example, for a candidate region of "100W straight fluorescent lamp" generated with a score of 90%, the model re-analyzes the features and finally classifies it as "100W straight fluorescent lamp". (2) On the other hand, the candidate range for "40W incandescent light bulb" is characterized by a rounded shape, so the model classifies it as a "40W incandescent light bulb" based on that characteristic. (3) A "20W circle tube" with multiple candidate regions is also ultimately identified as a "20W circle tube" because the shape of each candidate region is a round ring.

[0079] (4) Improving the accuracy of the bounding box Applying Non-Maximum Suppression (NMS) combines overlapping bounding boxes into a single region, improving accuracy. (1) If multiple bounding boxes are generated for the same "100W straight fluorescent lamp" symbol, NMS is used to keep only the bounding box with the highest score and delete the other duplicate boxes. (2) For example, if three bounding boxes with scores of 95%, 90%, and 85% are generated for the "100W straight fluorescent lamp" symbol, only the box with a score of 95% will be kept and the others will be discarded. (3) Ultimately, one bounding box remains for each symbol, thereby eliminating redundancy in the detection results.

[0080] (5) Final output The bounding box and position information of each symbol are output as precise coordinates on the architectural drawing and used for subsequent cost calculations and mapping. (1) As an example of output, a "100W straight fluorescent lamp" is displayed with coordinates (x1, y1, x2, y2) on the drawing, and a "40W incandescent light bulb" is also output along with its position coordinates. (2) This clarifies the location information of lighting fixtures, which can be smoothly reflected in cost calculations and proposals for LED lighting alternatives.

[0081] In this way, the FasterR-CNN model accurately detects equipment symbols in architectural drawings and performs a series of processes to determine their location and type. The information obtained in each step is passed on to the next process, enabling highly accurate analysis of equipment symbols.

[0082] 3. Overview of the overall system operation This system is an integrated platform that executes a series of processes, from analyzing architectural drawings to proposing optimal lighting fixture replacements, and even cost calculation and output. The following outlines the system's processing, from reading drawing data to outputting recognition results.

[0083] (1) Reading and preprocessing of drawing data The patching unit 8 performs format conversion and image correction of architectural drawings as follows, preparing a foundation for efficient operation in the next process. (1) Format conversion: Convert PDF files to PNG format. (2) Resolution standardization: Standardize low-resolution drawings to 300 DPI. (3) Noise reduction: Enhances lines and removes background noise. Specifically, for example, the patching unit 8 performs the following processing. (1) Read old architectural drawings (in PDF format) and convert them into standardized images. (2) Remove any faint smudges or lines that occurred during scanning, and make the symbols of the lighting fixtures clearly visible. (3) This improves the detection accuracy in the next process. (4) The pre-processed image data is passed to the detection unit 3 and used for analyzing the equipment symbols.

[0084] (2) Symbol detection and classification The detection unit 3 uses a deep learning model such as FasterR-CNN to perform the following analysis. (1) Region Proposal (RPN): Identify regions in the image that are likely to contain illumination symbols. (2) Symbol classification: Analyze the proposed domain and classify the symbols into specific categories (e.g., fluorescent tubes, incandescent light bulbs). (3) Assigning confidence levels: A confidence score is calculated for each category. Specifically, for example, the detection unit 3 performs the following processing: (1) Detect lighting fixtures in the drawing and obtain the following results. • Fluorescent tubes (100W): 50 locations, reliability 90% or higher. • Incandescent light bulbs (60W): 30 locations, reliability 85% or higher. (2) The coordinates of each symbol are output, and the detection results (symbol type and coordinate information) are sent to the mapping unit 6 and used for selecting a replacement.

[0085] (3) Symbol substitute mapping The mapping unit 6 performs the following processing by referring to the database unit 9 based on the symbol detection results. (1) Identifying the detection symbol: Identify the lighting fixture (e.g., 100W straight fluorescent lamp). (2) Selection of the optimal alternative: Match energy-saving fixtures such as "20W LED straight tube" from the database. (3) Obtain necessary specification information (price, power consumption, etc.). Specifically, for example, the detection unit 3 performs the following processing. (1) Map the 50 detected "100W straight fluorescent tubes" to "20W LED straight tubes" in the database. (2) Information on the selected replacement product (power consumption 20W, price ¥3,000) is obtained. (3) The mapping results are passed to the cost calculation module and used for detailed calculations.

[0086] (4) Cost calculation and evaluation of energy savings The calculation unit 7 performs the following calculations based on the information stored in the database unit 9. (1) Calculation of replacement costs: Add the equipment price and installation costs. (2) Calculation of energy reduction: The cost of reduction is calculated based on the difference in power consumption and annual usage time. (3) Calculation of the investment recovery period: Calculate the period until the initial costs are recovered through the savings. (4) Coordination with subsequent processes: The calculation results are reflected in reports and drawing data in the output unit. Specifically, for example, the calculation unit 7 performs the following processing: (1) When replacing a 100W fluorescent lamp with a 20W LED tube (a) Replacement cost: ¥3,000 (LED price) + ¥1,000 (installation fee) x 50 locations = ¥200,000. (b) Annual cost reduction: (100W - 20W) × 50 units × 3,000 hours × 0.10 / kWh = 120,000. (c) Payback period: ¥200,000 ÷ ¥120,000 = approximately 1.7 years.

[0087] (5) Output of results and generation of reports The output unit provides analysis results in both visual and data formats. (1) Visual highlighting: Highlight the symbol to be replaced on the drawing (e.g., red frame). (2) Report generation: Detailed results are output in CSV or JSON format. Specifically, for example, the output section performs the following processing: (1) On the drawing, the "100W fluorescent lamp" to be replaced is surrounded by a red frame and labeled with the annotation "20W LED straight tube (replacement cost ¥6,000)". (2) In JSON format output, it will be recorded as follows: "original_fixture":"100W straight fluorescent lamp" "led_fixture":"LED straight tube 20W" "exchange_cost":6000 "annual_saving":240 "payback_period":0.93

[0088] (6) Conclusion This system features highly integrated modules that work together seamlessly from drawing analysis to cost calculation and output. Through the following processes, it enables efficient and accurate proposals for architects and engineers. (1) The patching section 8 prepares the drawing data, (2) The detection unit 3 analyzes the symbol, (3) The mapping unit 6 selects an appropriate replacement, (4) The calculation unit 7 calculates the economic benefit, (5) The output unit provides the results in visual and data formats. This allows the system to function as one that maximizes the potential for energy conservation and cost reduction.

[0089] [1-3. Effects of the Embodiment] The effects of this embodiment are as follows:

[0090] (1) Deep learning-based object detection In the analysis of architectural drawings, pattern matching and rule-based methods have traditionally been widely used for symbol recognition. In contrast, this embodiment utilizes a FasterR-CNN model to learn the features of objects, including equipment symbols, through deep learning. This makes it possible to automatically identify a variety of equipment symbols on drawings with high accuracy without manually setting features or rules.

[0091] (2) Customization of specialized equipment symbol recognition through transfer learning This embodiment utilizes transfer learning to adapt a pre-trained object detection model to the recognition of equipment symbols on architectural drawings. This allows for efficient model customization without the need for large datasets. Conventional analysis methods rarely employ such adaptive techniques, and the combination of a dataset and training for identifying specialized equipment symbols is a novel approach.

[0092] (3) Improving classification accuracy through confidence scores In this embodiment, a reliability score is assigned to each symbol, and classification is performed based only on symbols with a high score, thereby improving the detection accuracy. This approach reduces the risk of misclassification, which was difficult with conventional simple recognition.

[0093] (4) LED Lighting Compatibility by Automatic Mapping Function This embodiment has a technology that automatically collates the detection results of lighting fixtures with the existing LED lighting database unit 9 and automatically determines the correspondence with new LED lighting fixtures. As a result, designers can quickly consider alternative candidates for each lighting fixture. This function is a unique technology not found in conventional architectural drawing analysis.

[0094] (5) Flexible Input and Diverse Data Augmentation This embodiment supports various input formats such as JPEG, PNG, and PDF, and also performs data augmentation such as rotation, inversion, and color adjustment, enabling it to handle drawings with different drawing styles. This flexibility greatly improves adaptability in the actual usage environment and enables support for various formats and drawing styles that were difficult for conventional systems to handle.

[0095] (6) Integration of Comprehensive Cost Calculation System In addition to the automated object detection function, this embodiment also integrates a cost calculation function based on the detection results. By collectively calculating the costs involved in installation and replacement, energy efficiency, etc. using the automatically detected lighting fixtures, it provides convenience not found in conventional analysis systems.

[0096] (7) Technically Most Excellent Points Particularly excellent points in this embodiment are the automation of specialized symbol recognition using deep learning and transfer learning, and the LED lighting compatibility by the automatic mapping function. While conventional architectural drawing analysis relied on manual feature setting and rule creation, this embodiment utilizes the self-learning ability of AI and can handle different symbols and diverse drawing styles, making it an innovative approach in architectural drawing analysis.

[0097] (8) Transfer Learning Section 4 Equipment symbols in architectural drawings have unique characteristics that differ from general object detection (e.g., differences in the shape of stairwell lights and fluorescent tubes). Since general object detection models cannot adequately capture these characteristics, it is necessary to build a symbol-specific model using transfer learning. By incorporating the transfer learning unit 4, it becomes possible to efficiently perform specialized learning for equipment symbols while utilizing basic features already learned in general object detection tasks. As a result, it can recognize a variety of lighting fixture symbols, such as 100W straight fluorescent tubes and circular tubes, with high accuracy.

[0098] (9) Accuracy improvement part 5 Because various factors (such as reduced resolution and noise) are involved in the detection of equipment symbols, a mechanism is needed to evaluate the reliability of the detection results and minimize misclassification. By incorporating the accuracy improvement unit 5, filtering based on the confidence score can be performed, eliminating misrecognitions with low confidence. Furthermore, by performing retraining on symbols that are frequently misclassified, the classification accuracy can be continuously improved. As a result, the ability to distinguish similar symbols, such as 100W straight fluorescent lamps and 20W straight fluorescent lamps, is improved.

[0099] (10) Mapping section 6 Not only is it essential to detect equipment symbols in architectural drawings, but also to map them to existing lighting fixtures and LED lighting fixtures, in order to concretize installation plans and energy-saving effects. By incorporating the mapping unit 6, it becomes possible to automatically match, for example, stairwell lights to energy-saving LED fixtures. As a result, designers can quickly select appropriate lighting fixtures, and the estimation of energy-saving effects and installation costs can be carried out efficiently.

[0100] (11) Patching unit 8 and detection unit 3 Architectural drawings come in a wide variety of formats and resolutions, and without a system that can process them uniformly, the performance of the model will be significantly affected. Furthermore, accurate detection of equipment symbols is fundamental to other functions of this invention (e.g., cost calculation and mapping). The inclusion of a patching unit 8 standardizes the input data, allowing the FasterR-CNN model to perform analysis in an optimal state. Furthermore, the detection unit 3 accurately detects and classifies 100W straight fluorescent tubes and circular tubes on architectural drawings, significantly improving the accuracy of subsequent processing.

[0101] (12) Database Section 9 To evaluate replacement costs and energy savings based on detected equipment symbols, detailed related data (e.g., installation costs, power consumption) is required. By incorporating the database unit 9, it becomes possible to quickly calculate, for example, the amount of power consumption reduction achieved by replacing a 40W incandescent light bulb with a 20W LED. Furthermore, by specifically showing replacement costs and energy-saving effects, it is possible to propose practical installation plans to customers and designers.

[0102] (13) Data output unit 10 It is necessary to provide information on detected equipment symbols in a format that can be easily used by designers and engineers. Furthermore, it is crucial to enable data integration with other systems. By including a data output unit 10, detection results on architectural drawings can be visually confirmed, and the data can also be provided as structured data in JSON or CSV format. As a result, visualization of detection data and use in other systems become easier, and LED lighting fixture replacement planning and installation work can proceed smoothly.

[0103] [2. Other Embodiments] The present invention is not limited to the embodiments described above, and the components can be modified and implemented in practice without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriate combinations of the multiple components disclosed in the above embodiments. For example, some components may be deleted from all the components shown in the embodiments. Moreover, the following embodiments are also included in the present invention.

[0104] (1) Simplification of the system Simplify the processing steps of the entire system and focus on the minimum necessary functions. (1) Omit the patching unit 8 When the input drawing is already in a standard format and resolution (e.g., JPEG format, 300 DPI), omit the patching unit 8 and directly input it to the detection unit 3. This is effective, for example, when dealing with already high-quality digital design drawings. (2) Integration of the database unit 9 and the mapping unit 6 Integrate the search function of the database unit 9 into the mapping unit 6 to directly select the corresponding LED lighting fixtures. This is suitable for small-scale facilities that handle only a small number of standard lighting fixtures.

[0105] (2) Advanced expansion example Add new functions to enhance the system. (1) Cooperation with environmental sensors Measure temperature, illuminance, power consumption, etc. in real time and optimize the proposal of lighting fixtures based on this. For example, systems for offices and schools that optimize not only energy savings but also illuminance. (2) Import of 3D design data Directly analyze not only 2D architectural drawings but also 3D BIM data (Building Information Modeling). It is applicable to design and analysis support systems for new buildings.

[0106] (3) Independence of partial functions Make a part of the system independent and divert it for other uses. (1) Use only the detection unit 3 Use the detection unit 3 alone to automatically extract only the equipment symbols from the architectural drawing. For example, connect and use it with an existing database unit 9 or mapping tool. (2) Use only the mapping unit 6 Use the mapping unit 6 alone to support the conversion to LED lighting fixtures based on existing lighting fixture information. For example, for facilities where a replacement plan has already been established.

[0107] (4) Expansion of the applicable range Expand the range of lighting fixtures and data formats covered. (1) Items other than lighting fixtures It detects not only lighting fixtures but also other equipment symbols such as air conditioners and ventilation fans. For example, it can be used for projects that examine the overall energy efficiency of equipment. (2) Handling of hand-drawn diagrams It can scan and recognize handwritten blueprints and old drawings. For example, it can be used for retrofitting projects.

[0108] (5) Automation of cost evaluation and design proposals The system will be given additional features such as design proposals and detailed cost analysis. (1) Proposal of multiple options It presents multiple LED fixture options and displays a comparison of their costs and energy-saving effects. For example, in a construction project where multiple design options are being considered. (2) Long-term cost assessment This calculates the lifecycle cost, including initial costs, maintenance costs, and operating costs. This is particularly useful for public facilities and schools with budget constraints.

[0109] (6) Adding or deleting parts It is also possible to transform it by freely combining the various parts. (1) Removal of the precision improvement unit 5 If a highly accurate model is available, the accuracy improvement unit 5 can be omitted. For example, if a highly reliable symbol classification model is already available. (2) Addition of new parts We will add an "effect prediction unit" that simulates the effects after replacing lighting fixtures. For example, a system that simulates energy efficiency and illuminance after replacement.

[0110] (7) Limitation of the target of detection Specializing for specific facilities or environments. (1) For schools: We specialize in classrooms and gymnasiums, offering solutions tailored to usage frequency and required illumination levels. For example, this is suitable for school facilities that primarily use straight fluorescent lights. (2) For hospitals: It can handle specialized lighting fixtures for operating rooms and patient rooms, and evaluates illuminance and color temperature. For example, it is suitable for supporting the planning of introducing medical LEDs. [Explanation of symbols]

[0111] 1…Dataset generation unit 2…Learning Department 3...Detection unit 4…Transfer Learning Department 5…Accuracy improvement part 6…Mapping section 7…Calculation section 8... Patching area 9…Database 10...Data output section

Claims

1. A drawing analysis system for detecting and classifying equipment symbols contained in a drawing, A dataset generation unit that generates multiple types of graphic datasets, including equipment symbols, A learning unit that uses a deep learning model to learn the features of equipment symbols based on the aforementioned dataset, A detection unit that applies the deep learning model to an unknown dataset and automatically detects and classifies multiple types of equipment symbols, A drawing analysis system equipped with the following features.

2. The drawing analysis system according to claim 1, comprising a transfer learning unit that uses a general object detection model as a pre-trained model and enables feature extraction specific to equipment symbols in the drawing through transfer learning.

3. The drawing analysis system according to claim 1 or 2, further comprising: a confidence score assignment unit that assigns a confidence score to the detection results output by the deep learning model; and an accuracy improvement unit that improves the classification accuracy of equipment symbols based on the confidence score.

4. A drawing analysis system according to claim 1 or 2, comprising: an identification unit that identifies existing lighting fixtures based on the detection results of equipment symbols; and a mapping unit that automatically searches a predefined database for the type of new LED lighting fixture corresponding to the detection results and maps each lighting fixture to an LED lighting fixture.

5. A drawing analysis system according to claim 1, comprising: an interface for accepting image files in JPEG, PNG, and PDF formats as drawing data; a patching unit for converting the accepted drawing data into a standardized input format according to its size and resolution; and a detection unit for detecting and classifying equipment symbols based on the standardized drawing data.

6. The drawing analysis system according to claim 1, comprising: a cost calculation unit that calculates the cost of replacing lighting fixtures and the amount of energy saved based on the drawing analysis results; and a database unit that references data on the installation cost, power cost, and replacement cost of LED lighting fixtures corresponding to the detected equipment symbols.

7. The drawing analysis system according to claim 1, comprising: an output unit that outputs detected equipment symbols in visual and structured data formats; an overlay unit that highlights the location of the detected equipment symbols on the drawing; and a data output unit that outputs the detection results in JSON or CSV format, enabling data integration with other systems.

Images