TRAINING FACILITY, TRAINING SYSTEM, MEDIUM AND INFORMATION PROCESSING METHOD FOR TRAINING FACILITY

Abstract

A training device according to the technology disclosed herein comprises: a training image acquisition unit (21) that acquires a training image; a feature extraction unit (22) that computes a common feature space feature of the training image; an existing feature acquisition unit (23) that acquires a prestored trained model and an existing feature corresponding to the prestored trained model; a feature comparison unit (24) that computes a distance or similarity between the common feature space feature and the existing feature as an index in a common feature space; a model selection unit (25) that, as a base model, selects one of the trained models suitable for a training purpose based on the index; and a model training unit (26) that performs retraining for the base model.a model evaluation unit (27) that evaluates the inference performance of the newly trained base model; and a trained model output unit (28) that outputs the newly trained base model.

Description

TECHNICAL FIELD

[0001] The technology disclosed herein relates to a training facility and a training system. BACKGROUND ON THE STATE OF THE TECHNOLOGY

[0002] In the technical field of machine learning, a method for increasing development efficiency is known in which an existing training model is successfully used instead of creating it from scratch.

[0003] Patent literature 1, for example, describes an information processing device that uses a method to select a model from a multitude of models. This model is trained on data corresponding to an attribute of a detection target and exhibits good detection performance for that target. Patent literature 1, for instance, cites the data attribute "a type of infrastructure structure such as a bridge or a tunnel" as an example when training a model that detects a crack.

[0004] The data attribute assumed in patent literature 1 is defined by a human being who is a user and is based on a human concept such as "bridge", "tunnel" or "dam". REFERENCE LISTPATENT LITERATURE

[0005] Patent Literature 1: JP 2021-81793 A SUMMARY OF THE INVENTIONTECHNICAL PROBLEM

[0006] Although patent literature 1 focuses on the data attribute, it is not possible to quantitatively determine whether a training model is suitable for a training purpose based solely on the data attribute. For example, in a case where there is a training model a and a training model b, both trained on training datasets with the same data attributes such as "bridge," "tunnel," and "dam," but with different images, it is impossible to know which training model to use. In the prior art, it is necessary in such a case to perform a recognition performance test of both training model a and training model b. This does not significantly increase development efficiency.

[0007] One objective of the technique disclosed herein is to provide a training facility and training system that introduces an index that quantitatively indicates whether an existing training model is suitable for a training purpose or not, and that has a higher development efficiency than related techniques. SOLUTION TO THE PROBLEM

[0008] A training facility according to the present disclosed technique comprises: a training image acquisition unit to acquire a training image; a feature extraction unit to compute a common feature space feature of the training image; an existing feature acquisition unit to acquire a pre-stored trained model and an existing feature corresponding to the pre-stored trained model; a feature comparison unit to compute, in a common feature space, a distance or similarity between the common feature space feature and the existing feature as an index; a model selection unit to select, as a base model, one of the trained models suitable for a training purpose based on the index; and a model training unit to perform retraining on the base model.a model evaluation unit to assess the inference performance of the newly trained base model; and a trained model output unit to output the newly trained base model. ADVANTAGEOUS EFFECTS OF THE INVENTION

[0009] The training facility according to the technique disclosed herein has the configuration described above and can therefore quantitatively indicate whether an existing training model is suitable for a training purpose or not, and has a higher development efficiency than related techniques. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is an explanatory illustration which depicts a configuration of a training system according to a first embodiment. Fig. Figure 2 is a block diagram representing a functional configuration of a training facility according to the first embodiment. Fig. 3A and Fig. 3B are explanatory illustrations that depict a hardware configuration of the training facility according to the first embodiment. Fig. Figure 4 is a flowchart that depicts the processing steps of the training facility according to the first embodiment. DESCRIPTION OF THE EXECUTION FORMS

[0010] It can be stated that a problem solved by a training system 1 according to the technique disclosed herein is a specific type of optimization problem, similar to a general machine learning problem. A no-free-lunch theorem is known in relation to this optimization problem. The no-free-lunch theorem states, in short, that a model with high versatility and high performance is theoretically impossible. One model (MA) performs better than another model (MB) because the model (MA) is engineered or specialized for a particular problem to be solved.

[0011] It can be stated that a feature of the technology disclosed herein is that a model that has been trained in a general or universal way over a wide area and a model that has been trained in a specialized way in a particular area can be used selectively with inventive effort. First embodiment.

[0012] Fig. Figure 1 is an explanatory illustration depicting a configuration of a training system 1 according to a first embodiment. As shown in Fig. As shown in Figure 1, the training system 1 according to the first embodiment comprises a training device 2, an operation input device 3, a storage device 4 and a display output device 5.

[0013] As in Fig. As shown in Figure 1, the training facility 2, the operation input facility 3, the storage facility 4 and the display output facility 5 can be interconnected via an information network such as a local area network (LAN). <<Trainingseinrichtung 2 als Bestandteil des Trainingssystems 1> >

[0014] The training facility 2, forming the training system 1, is a facility with a mathematical model (also referred to as the "training model") that is capable of performing supervised training based on a training dataset containing a training image or the like.

[0015] Details of training facility 2 can be found in the following description. <<Operationseingabeeinrichtung 3 als Bestandteil des Trainingssystems 1> >

[0016] The operating input device 3, which forms the training system 1, is a user interface that receives a user input (hereinafter simply referred to as "user input") and outputs a corresponding operating signal. The operating input device 3 is a device such as a keyboard, a pointing device, a touch panel, or a touch sensor.

[0017] The operational signal output by the operational input device 3 is sent to the training device 2. <<Speichereinrichtung 4 als Bestandteil des Trainingssystems 1> >

[0018] Storage device 4, which forms training system 1, is a device that stores information required for the processing carried out by training device 2. As in Fig. As shown in Figure 1, the storage device 4 is located outside the training device 2. Inside the training device 2 is a working memory 7, which will be described later, and in this description the working memory 7 and the storage device 4 are distinguished from each other.

[0019] As in Fig. As shown in Figure 1, the training facility 2 can access the storage device 4. Therefore, the training facility 2 can read information stored in the storage device 4 and store information in the storage device 4. <<Anzeigeausgabeeinrichtung 5 als Bestandteil des Trainingssystems 1> >

[0020] The display output device 5, which forms the training system 1, is a user interface that transmits visual information to a user. Display output device 5 is, for example, a display. The visual information transmitted to the user is transferred from the training device 2.

[0021] Fig. Figure 2 is a block diagram showing a functional configuration of the training facility 2 according to the first embodiment. As in Fig. As shown in Figure 2, the training facility 2 comprises a training image acquisition unit 21, a feature extraction unit 22, an existing feature acquisition unit 23, a feature comparison unit 24, a model selection unit 25, a model training unit 26, a model evaluation unit 27, and a trained model output unit 28. <<Trainingsbild-Erwerbungseinheit 21 in Trainingseinrichtung 2> >

[0022] The training image acquisition unit 21 in the training facility 2 is a component that acquires a training data set, for example, a training image.

[0023] Generally, there are a large number of training images. It is assumed that the training image is stored, for example, in memory unit 4. The training image acquisition unit 21 acquires the training image stored in memory unit 4.

[0024] The training image acquired by the training image acquisition unit 21 is transferred to the feature extraction unit 22. <<Merkmalsextrahierungseinheit 22 in Trainingseinrichtung 2> >

[0025] The feature extraction unit 22 in training facility 2 is a component that extracts a feature from the transferred training image. In this case, a feature is a term commonly used in the technical field of machine learning and, in short, refers to numerical data that expresses a characteristic such as a thing or an event. More specifically, a feature is a characteristic or attribute of data that can be used by a model to make predictions.

[0026] One role of the feature extraction unit 22 is to function as an AI facility that has undergone preparatory training in image recognition; that is, it has acquired in advance, through training, a method for converting target data into an expression that can easily solve a problem. The expression method acquired through this expression training is the feature.

[0027] The feature extracted by feature extraction unit 22 is referred to in this description as the "common feature space feature" to distinguish it from a feature used as a generic term. The common feature space feature can also be viewed as a vector defined in the "common feature space." More generally, when specific image data is fed into a training device, a feature obtained as an intermediate product can vary depending on the training process (progress). On the other hand, the common feature space feature defined by the technique disclosed herein does not vary as an expression of specific image data.

[0028] The common feature space is a large-scale map, such as a world map, to use a figurative or visual analogy. A world map is suitable for gaining an overview, as it can contain a wide range of information, but it is not suitable for verifying detailed information. Details of how the common feature space is used in the technology disclosed herein will follow from the description below.

[0029] Feature extraction unit 22 comprises a mathematical model for extracting a common-feature-space feature. A well-known mathematical model is, for example, an artificial neural network (also simply referred to as a "neural network").

[0030] In the training facility 2 according to the present disclosed technique, the mathematical model contained in the feature extraction unit 22 is preferably ResNet, which is trained by ImageNet, a general image dataset. The following document serves as a reference for ResNet.

[0031] “Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun,” “Deep Residual Learning for Image Recognition,” “Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016 pp. 770-778.”

[0032] Generally, the feature comprises a local feature and a global feature in an image. It is also considered that the local feature could, for example, include a characteristic related to an edge and a characteristic related to a sub-shape. It is also conceivable that the global feature could, for example, contain a characteristic specific to a category to which a teacher label belongs.

[0033] A sufficiently trained training model (referred to as a "trained model") extracts a feature suitable for achieving a training purpose. As described above, what the trained model extracts as a feature is necessary to achieve a training purpose and therefore does not necessarily correspond to a feature of an image viewed by a human, for example, "with a texture like metal," "with a triangular shape as a whole," or "a dog appears."

[0034] The reason for this is that "metal," "texture," "triangle," "dog," and similar terms are human-created, while the mathematical model does not recognize these concepts. The artificial intelligence, which uses a trained model, is only able to extract a single feature suitable for achieving a training goal through machine learning.

[0035] Feature Extraction Unit 22 can use a method called Bag-of-Visual-Words (BOW) to extract a Common Feature Space feature. BOW is a method for dimensionally compressing an image into a one-dimensional vector (also called a "feature vector"). BOW can also be referred to as Bag-of-Words or Bag-of-Features. The following document serves as a reference for BOW.

[0036] “Csurk, Gabriella, et al.,” “Visual categorization with bags of keypoints.”, Workshop on statistical learning in computer vision, ECCV. Edition 1. No. 1-22. 2004.

[0037] As described above, a well-known training model can be used as the feature extraction unit 22. Recently, applications that can perform a search simply by taking an image or holding a camera have become widespread, and the feature extraction unit 22 can employ a technique used in such an application. Furthermore, the feature extraction unit 22 can utilize a well-known feature extraction method.

[0038] Since the common feature space, as described above, functions as a large-scale map, such as a world map, it is preferable to use a training model with many classification classes that has undergone preliminary training with a broad range of general images to ensure versatility. Furthermore, it is desirable that the common feature space feature does not vary as an expression of specific image data. Therefore, the training model for feature extraction unit 22, used to extract the common feature space feature, is not further trained but employed with a fixed parameter.

[0039] The common feature space feature extracted by the feature extraction unit 22 is transferred to the feature comparison unit 24 together with an original training image. <<Vorhandenes-Merkmal-Erwerbungseinheit 23 in Trainingseinrichtung 2> >

[0040] The existing feature acquisition unit 23 in the training facility 2 is a component that acquires "existing features" which were previously stored in the storage facility 4.

[0041] The training system 1 according to the presently disclosed technique contains several types of pre-trained models in the storage device 4. These trained models are pre-stored in the storage device 4 as candidates for a training model, which is selected according to the purpose of the training. In contrast to a training model trained for a general purpose and contained in the feature extraction unit 22, the trained model to be stored in the storage device 4 is a specialized model trained for a specific purpose. Although the trained model stored in the storage device 4 is literally a previously trained model, the training is continued in the model training unit 26 described later (referred to as "retraining"). Details of the retraining will follow from the description below.

[0042] The training system 1 according to the technique disclosed herein adds a specific index to the trained model, which is stored in the memory device 4, and a basis of the index is an existing feature.

[0043] For example, it is assumed that a trained model (MP), to be stored in memory facility 4, was trained while specializing in image recognition of a flower. When image data is input into the trained model (MP), the trained model (MP) can also generate a feature as an intermediate product. A feature space in the trained model (MP) is a detailed map only for a specific region, using a figurative expression. The feature generated by the trained model (MP) is a quantity that is meaningful only within a feature space defined only in the trained model (MP) and cannot be used as an index for comparison with another model.

[0044] For example, it is assumed that the trained model (MP) was trained such that "spring flowers => 1, summer flowers => 2, autumn flowers => 3, and winter flowers => 4". It is further assumed that another trained model (MV) was trained and simultaneously specialized for a vertebrate such that "dog => 1, cat => 2, and others => 3". In general, a feature vector obtained in this way as an intermediate product is not uniform with respect to the meaning of any of its elements and any of its dimensions.

[0045] The technique disclosed herein focuses primarily on the type of training dataset used for training the trained model (MP). In this technique, a common feature space feature is calculated for each of the images used for training the MP. This common feature space feature is calculated in advance using a method similar to that used in the feature extraction unit 22. A representative point of this common feature space feature is stored in the storage device 4 as an existing feature. By calculating the common feature space feature in this way, the trained model (MP) can be expressed as a model in which a common feature space feature of an image used for training is plotted in the common feature space.When using a figurative expression, the trained model (MP) can be expressed as a plot distributed over an area corresponding to the “flower” in a common feature space that is a world map.

[0046] In the technique disclosed herein, for example, a representative point such as the centroid of a common feature space feature is assumed, which is plotted in a common feature space as an existing feature to determine an index of the trained model (MP). Using a figurative expression, it can be stated that the existing feature of the trained model (MP) is a center of the area of ​​the "flower" in the common feature space, which is a world map.

[0047] Besides a flower, a trained model stored in memory location 4 could be an insect, a bird, an animal, a tall person, a national flag, a work of art, or a model specializing in any other picture book theme. Examples of trained models stored in memory location 4 also include a model specializing in an unoccupied register, a model specializing in character recognition such as optical character recognition (OCR), a model specializing in biometric authentication, a model specializing in three-dimensional shape recognition, a model specializing in robot control, a model specializing in autonomous driving, a model specializing in crime prevention, and a model specializing in various other areas.

[0048] The candidate of the training model and the existing feature acquired by the existing feature acquisition unit 23 are transferred to the feature comparison unit 24. <<Merkmalsvergleichseinheit 24 in Trainingseinrichtung 2> >

[0049] The feature comparison unit 24 in training facility 2 is a component for comparing a common feature space feature within a common feature space. More precisely, the processing performed by feature comparison unit 24 consists of calculating the distance between two or more common feature space features (vectors) plotted in the common feature space.

[0050] In other words, if the distance between two common feature space features in the common feature space is small, this means that the two original images have similar characteristics or attributes.

[0051] A starting point of a distance in the common feature space, calculated by the feature comparison unit 24, is the common feature space feature extracted by the feature extraction unit 22. An endpoint of the distance calculated by the feature comparison unit 24 is a plurality of existing features corresponding to the plurality of trained models, each acquired by the existing feature acquisition unit 23.

[0052] Since, as described above, a large number of training images can be entered into the training image acquisition unit 21, the common feature space feature, which refers to the training image plotted in the common feature space, is distributed. Therefore, the feature comparison unit 24 can use a Mahalanobis distance, commonly used in statistics, as the distance in the common feature space.

[0053] The core of the processing performed by the feature comparison unit 24 is to search for a trained model that was trained on image data exhibiting similar features or attributes to an input training image. The value calculated by the feature comparison unit 24 is not limited to a "distance" such as the Euclidean distance or the Mahalanobis distance from such a viewpoint. The feature comparison unit 24 can calculate a value using a similarity concept, such as cosine similarity.

[0054] The value (distance, similarity, etc.) calculated by the feature comparison unit 24 for each trained model is transferred to the model selection unit 25 as an index for selecting a trained model that is suitable for a training purpose. <<Modellauswahleinheit 25 in Trainingseinrichtung 2> >

[0055] The model selection unit 25 in the training facility 2 is a component that directly selects a trained model that is best suited for a training purpose, or suggests a candidate for a trained model to be selected to a user based on the index transmitted by the feature comparison unit 24.

[0056] In a case where the determination of a user is not necessary, the model selection unit 25 directly selects a trained model with the smallest distance in the common feature space or the highest similarity based on the index transmitted by the feature comparison unit 24.

[0057] In a case where the trained model to be used is determined by a user, the model selection unit 25 lists the trained models, for example in ascending order of distance in the common feature space or in descending order of similarity, and displays the listed trained models on the display output device 5. The user selects the trained model to be used from the list displayed on the display output device 5 using the operation input device 3.

[0058] The trained model selected directly or indirectly by the model selection unit 25 is transmitted as the base model to the model training unit 26. <<Modelltrainingseinheit 26 in Trainingseinrichtung 2> >

[0059] Model training unit 26 in training facility 2 is a component that performs retraining for the base model selected by model selection unit 25. The training dataset for model training unit 26 to perform retraining for the base model can be a training image acquired via training image acquisition unit 21.

[0060] As described above, the training image for calculating the common feature space feature in feature extraction unit 22 can be one or a multitude of training images acquired via training image acquisition unit 21. Conversely, model training unit 26 uses, in principle, all but one remaining training image for evaluation from among those acquired via training image acquisition unit 21 for retraining.

[0061] The newly trained base model and the training image remaining for evaluation are transmitted to the model evaluation unit 27. <<Modellbewertungseinheit 27 in Trainingseinrichtung 2> >

[0062] The model evaluation unit 27 in the training facility 2 is a component that evaluates an inference performance of the newly trained base model using the training image remaining for evaluation.

[0063] An evaluation result performed by the model evaluation unit 27 is preferably displayed on the display output device 5 to notify the user of the evaluation result. It should be noted that, according to the present disclosed technique, the training system 1 can be configured to display the evaluation of the training performance on the display output device 5 together with the evaluation of the inference performance.

[0064] Based on the evaluation result displayed on the display output device 5, the user determines whether to complete the retraining or to continue the retraining by adding a training data set.

[0065] In a case where it is determined that retraining is complete, the newly trained base model is transmitted to the Trained Model Output Unit 28. <<Trainiertes-Modell-Ausgabeeinheit 28 in Trainingseinrichtung 2> >

[0066] The trained model output unit 28 in the training facility 2 is a component that stores the newly trained base model in the storage facility 4.

[0067] Fig. Figure 3 is an explanatory illustration showing a hardware configuration of the training facility 2 according to the first embodiment.

[0068] In the training device 2 according to the first embodiment, the functions of the training image acquisition unit 21, the feature extraction unit 22, the existing feature acquisition unit 23, the feature comparison unit 24, the model selection unit 25, the model training unit 26, the model evaluation unit 27, and the trained model output unit 28 are implemented by a processing circuit. The processing circuit can either be a processor 6 that executes a program stored in the main memory 7 (see Fig. 3A) or a processing circuit 8, which is dedicated hardware (see Fig. 3B).

[0069] Fig. 3A is a hardware configuration representation in a case where the functions of training facility 2 are executed by software. In the case of a Fig. In the hardware configuration shown in Figure 3A, the functions of the training device 2 are implemented by software, firmware, or a combination of both. The software and the firmware are each described as a program and stored in the main memory 7. By reading and executing a program stored in the main memory 7, the processor 6 (also referred to as a CPU, central processing unit, processing unit, microprocessor, microcomputer, or DSP) implements the function of each unit. That is, the training device 2 includes the main memory 7 for storing a program that, when executed by the processor 6, causes the functions of the training device 2 to be output.Furthermore, it can also be determined that these programs cause a processor 6 to execute a sequence, method, or procedure performed by the training device 2. The working memory 7 can be non-volatile or volatile semiconductor memory such as RAM, ROM, flash memory, or EPROM. In addition, the working memory 7 can be a disk such as a magnetic disk, flexible disk, optical disk, compact disk, mini-disk, or DVD. Furthermore, the working memory 7 can be in the form of an HDD or SSD.

[0070] Fig. 3B is a hardware configuration representation in a case where the functions of training facility 2 are performed by dedicated hardware. In the case of the in Fig. The hardware configuration shown in Figure 2B corresponds, for example, to a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof, to the processing circuit 8. The functions of the training device 2 can each be implemented by individual processing circuits 8 or jointly by a processing circuit 8.

[0071] It should be noted that some of the functions of Training Facility 2 may be implemented through software or firmware, while the remaining functions may be implemented through dedicated hardware. Thus, Training Facility 2 can implement its functions through hardware, software, firmware, or a combination thereof.

[0072] Fig. Figure 4 is a flowchart showing the processing steps of the training facility 2 according to the first embodiment. As in Fig.As shown in Figure 4, the processing steps of the training device 2 according to the first embodiment comprise a processing step of acquiring the training image (ST201), a processing step of extracting a feature (ST202), a processing step of acquiring an existing feature (ST203), a processing step of calculating a similarity (ST204), a processing step of selecting a model (ST205), a processing step of performing model retraining (ST206), a processing step of evaluating a model (ST207), and a processing step of outputting information about the trained model (ST208).

[0073] The processing step of acquiring a training image (ST201) is a processing step that is performed by the training image acquisition unit 21.

[0074] The processing step of extracting a feature (ST202) is a processing step that is performed by the feature extraction unit 22.

[0075] The processing step of acquiring an existing feature (ST203) is a processing step that is carried out by the existing feature acquisition unit 23.

[0076] The processing step of calculating a similarity (ST204) is a processing step that is performed by the feature comparison unit 24.

[0077] The processing step of selecting a model (ST205) is a processing step that is performed by the model selection unit 25.

[0078] The processing step of performing model retraining (ST206) is a processing step performed by the model training unit 26.

[0079] The processing step of evaluating the model (ST207) is a processing step that is performed by the model evaluation unit 27.

[0080] The processing step of outputting information about the trained model (ST208) is a processing step performed by the Trained Model Output Unit 28.

[0081] One of the notable effects of training system 1 according to the technique disclosed herein is that a model that has been trained in a general way over a wide range and a model that has been trained in a specialized way in a specific range can be used selectively.

[0082] One of the most noteworthy effects of the training system 1 according to the technique disclosed herein is that an index (distance or similarity) that quantitatively indicates whether an existing training model is suitable for a training purpose or not can be proposed by defining a common feature space.

[0083] One of the most notable effects of the training system 1 according to the technique disclosed herein is that the number of times an unnecessary recognition performance test is performed is reduced and development efficiency is improved compared to the state of the art. INDUSTRIAL APPLICABILITY

[0084] The training system 1 and the training facility 2) according to the technology disclosed herein can be widely used in the technical field of image recognition and are industrially applicable. REFERENCE MARK LIST

[0085] 1: Training system, 2: Training device, 3: Operation input device, 4: Memory device, 5: Display output device, 6: Processor, 7: Main memory, 8: Processing circuit, 21: Training image acquisition unit, 22: Feature extraction unit, 23: Existing feature acquisition unit, 24: Feature comparison unit, 25: Model selection unit, 26: Model training unit, 27: Model evaluation unit, 28: Trained model output unit QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] JP 2021-81793 A

[0005] Cited non-patent literature

[0000] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun“, „Deep Residual Learning for Image Recognition“, „Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016 pp. 770-778

[0031] Csurk, Gabriella, et al.“, „Visual categorization with bags of keypoints.“, Workshop on statistical learning in computer vision, ECCV. Aufl. 1. Nr. 1-22. 2004

[0036]

Claims

[1] Training facility comprising: a training image acquisition unit for acquiring a training image; a feature extraction unit to calculate a common feature space feature of the training image; an existing feature acquisition unit for acquiring a pre-stored trained model and an existing feature corresponding to the pre-stored trained model; a feature comparison unit for calculating, in a common feature space, a distance or similarity between the common feature space feature and the existing feature as an index; a model selection unit for selecting, as a base model, one of the trained models suitable for a training purpose based on the index; a model training unit to perform retraining for the base model; a model evaluation unit to evaluate an inference performance of the newly trained baseline model; and a trained model output unit to output the newly trained base model. [2] A training system comprising a training device, an operation input device, a storage device and a display output device connected to each other, the training device comprising: a training image acquisition unit for acquiring a training image; a feature extraction unit to calculate a common feature space feature of the training image; an existing feature acquisition unit for acquiring a pre-stored trained model and an existing feature corresponding to the pre-stored trained model; a feature comparison unit for calculating, in a common feature space, a distance or similarity between the common feature space feature and the existing feature as an index; a model selection unit for selecting, as a base model, one of the trained models suitable for a training purpose based on the index; a model training unit to perform retraining for the base model; a model evaluation unit to evaluate an inference performance of the newly trained baseline model; and a trained model output unit to output the newly trained base model.

Images