Learning Device for Multi-Object Classification via Region Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image classification methods using convolutional neural networks require a large number of images for accurate learning and struggle to classify multiple objects within a single image effectively.
Innovation Solution
A learning device that generates a model with a feature extraction part and a coordinate estimation part, which extracts and further trains on images of specific object regions, increasing the dataset by automatically generating more images with varying object positions and ratios, and includes class estimation parts for improved classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of images are prepared for learning, then classification accuracy is improved, but it becomes difficult to prepare sufficient images in advance
Solution Approach 1:
The patent creates synthetic training images by copying and recombining object regions from existing images. The system extracts object regions from source images, generates new background images, and composes synthetic images by placing extracted objects onto generated backgrounds. This copying approach enables generation of large numbers of training images without requiring manual collection of additional real images.
Solution Approach 2:
The patent performs preliminary extraction of object regions and their features from source images before the actual training process. By pre-processing and storing object region data, the system prepares training materials in advance, enabling efficient generation of synthetic training images during the learning process without time-consuming real-time processing.
2Adaptability or versatility
If conventional classification methods are used, then processing is simpler, but multiple objects within a single image cannot be classified effectively
Solution Approach 1:
The patent segments images into multiple object regions and processes each region independently through the convolutional neural network. The system divides a single image containing multiple objects into separate region-of-interest images, allowing the model to classify each object individually. This segmentation approach enables multi-object classification while maintaining a relatively simple base model structure.
Solution Approach 2:
The patent transforms the classification problem from image-level to region-level by extracting and processing object regions as separate entities. Instead of classifying entire images, the system operates on extracted object regions, adding a spatial dimension to the classification process. This allows simultaneous classification of multiple objects within a single image by treating each object region as an independent classification target.
Data Source
AI summary
A learning device acquires a plurality of images from a storage. The learning device trains a feature extraction part using a plurality of images. The learning device causes the trained feature extraction part to output a feature map of each of one or more images among the plurality of images. The learning device inputs the output feature map to a coordinate estimation part for each of one or more images, and causes the coordinate estimation part to output at least one set of coordinates of a region of a portion where the object is present in the image. The learning device extracts an image of a portion indicated by the output coordinates from the image, for each of one or more images. The learning device further trains the trained feature extraction part using the extracted image.