Object Recognition via Component Segmentation and Combination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition methods, such as SSD, face challenges in accurately recognizing regular constituent objects from images, especially when training images are tilted or contain components with varying sizes, colors, and shapes, leading to incorrect combinations and increased erroneous recognitions due to the need for large amounts of training data and the complexity of combinatorial optimization.
Innovation Solution
An estimation program and apparatus that utilize a learned detection model to identify primary and secondary components within an image, model their relationships, and estimate candidate components, allowing for accurate recognition of regular constituent objects without requiring extensive training data or solving the combinatorial optimization problem for all components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large amount of training data is used to recognize components with varying sizes, colors, and shapes, then recognition coverage is improved, but learning difficulty and computational complexity increase significantly
Solution Approach 1:
The patent segments the recognition task into two distinct stages: (1) component detection using a pre-trained detection model to identify individual components and their attributes, and (2) combination determination using a separate combination model to identify which components form regular constituent objects. This segmentation allows each model to be specialized and trained independently, reducing the overall learning complexity while maintaining comprehensive recognition coverage.
Solution Approach 2:
The patent applies preliminary action by pre-training the detection model on a large dataset of individual components with varying attributes (sizes, colors, shapes) before using it in the combination stage. This preliminary training ensures the detection model has already learned to recognize diverse component variations, eliminating the need to retrain on all possible combinations during the final recognition task.
2Measurement precision
If all possible combinations of components are trained to ensure accurate recognition, then recognition accuracy is improved, but the number of training samples required becomes prohibitively large
Solution Approach 1:
The patent divides the training data requirement into two separate, manageable portions: (1) training data for the detection model to recognize individual components and their attributes, and (2) training data for the combination model to learn spatial relationships and combination patterns. This segmentation eliminates the need to create and train on all possible component combinations, dramatically reducing the total training data volume while maintaining high recognition accuracy.
Solution Approach 2:
The patent introduces an intermediary detection model that acts as a bridge between the input image and the combination recognition task. This detection model first identifies and extracts individual components with their attributes, providing cleaned and structured input to the combination model. This intermediary step reduces the complexity and volume of training data needed for the final combination recognition task.
3Device complexity
If component detection is performed before combination determination, then processing structure is simplified, but erroneous recognition increases when training images are tilted or misaligned
Solution Approach 1:
The patent applies parameter changes by transforming the detection and combination tasks into different parameter spaces. The detection model operates in component attribute space (identifying what components are present and their properties), while the combination model operates in spatial relationship space (determining how components are positioned and combined). This parameter transformation allows each model to focus on specific aspects, improving reliability while maintaining processing simplicity.
4Measurement precision
If a detection model is trained to identify individual components, then component detection capability is improved, but the ability to correctly determine combinations of components deteriorates
Solution Approach 1:
The patent segments the overall recognition system into two specialized models: a detection model optimized for identifying individual components with high precision, and a combination model optimized for determining which detected components form regular constituent objects. This segmentation allows each model to excel at its specific task without compromising the other, achieving both high detection precision and accurate combination recognition.
Solution Approach 2:
The detection model serves multiple functions: it not only detects the presence of components but also extracts their attributes (size, color, shape, position) which are then fed to the combination model. This multi-functionality of the detection model provides comprehensive input data that enables the combination model to accurately determine object compositions without requiring the detection model itself to be retrained for combination tasks.
Data Source
AI summary
An estimation program causing a computer to execute a step (S104) of acquiring a learned detection model and a constituent component model, a step (S202) of outputting constituent component detection results based on the new image using the learned detection model, a step (S203) of estimating first primary component candidates based on the constituent component model and the constituent component detection results and sorting out a secondary component not registered with a list of the first primary component candidates, a step (S204) of merging a plurality of the first primary component candidates to generate a second primary component candidate, and a step (S205) of selecting a secondary component matching the second primary component candidate from among a plurality of secondary components sorted out and registering the selected secondary component with the second primary component candidate.