Object Recognition Model Training with Unregistered Category Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition techniques face challenges in accurately distinguishing between registered and unregistered category objects during the identification process, as the identification score for unregistered category objects can be higher than that of registered category objects, leading to incorrect rejection or acceptance of identification results.
Innovation Solution
A learning device and method that acquire both first training data for registered category objects and second training data for unregistered category objects, calculate identification scores and unregistered scores, and learn an identification model to differentiate between registered and unregistered categories based on these scores, ensuring accurate identification and rejection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If only first training data regarding registered category objects is used for learning, then the identification model can be trained with simpler data, but the identification score for unregistered category objects becomes higher than that for registered category objects, leading to incorrect identification results
Solution Approach 1:
The patent applies preliminary action by pre-collecting and storing second training data regarding unregistered category objects before the actual identification process. This preparatory step ensures that when unregistered objects are encountered during identification, the model has already been exposed to such data during training, enabling it to correctly distinguish them and assign appropriate low identification scores, thereby preventing incorrect identification results.
Solution Approach 2:
The patent segments the training data into two distinct categories: first training data regarding registered category objects and second training data regarding unregistered category objects. This segmentation allows the model to learn different characteristics separately - registered objects for positive identification and unregistered objects for rejection - thereby resolving the contradiction between training simplicity and identification accuracy.
2Measurement precision
If the identification model is trained using both first training data and second training data, then the identification accuracy improves, but the data acquisition and processing complexity increases
Solution Approach 1:
The patent makes the training system universal by designing a unified learning process that handles both registered and unregistered category objects through the same neural network architecture. The identification model performs multiple functions: it identifies registered categories while simultaneously recognizing and rejecting unregistered categories. This multi-functionality reduces overall system complexity compared to maintaining separate models for each task.
3Reliability
If second training data regarding unregistered category objects is acquired and processed, then the model can correctly reject unregistered objects, but the learning time and computational resources increase
Solution Approach 1:
The patent applies partial action by selectively processing only the necessary portions of training data. The system acquires second training data regarding unregistered category objects but processes it through efficient sampling and batch training approaches. This allows the model to learn the characteristics of unregistered objects sufficiently for accurate rejection without undergoing exhaustive training on all possible unregistered object variations, thereby reducing learning time while maintaining reliability.
Data Source
AI summary
A learning device is configured to acquire first training data that is training data regarding a registered category object and second training data regarding an unregistered category object, the registered category object belonging to a registered category which is registered as an identification target, the unregistered category object not belonging to the registered category. The learning device is also configured to calculate, on a basis of the first training data and the second training data, an identification score indicating a degree of certainty that an object to be identified belongs to the registered category. The learning device is also configured to calculate, on a basis of the second training data, an unregistered score indicating a degree of certainty that the object to be identified does not belong to the registered category. The learning device is also configured to learn an identification model which performs an identification regarding the registered category based on the identification score and the unregistered score.