Deep Neural Network Sub-object Region Picking for Image Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image analysis methods using HA-CNN struggle with imperfect cut-out images, where the order of sub-objects is different or only a part of the object is reflected, leading to reduced analysis precision.
Innovation Solution
A learning method that allocates feature values of sub-objects to specific modules in a deep neural network, performing direct and indirect learning to precisely pick up regions of sub-objects, using information about the regions in images and improving analysis precision based on extracted feature values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the feature value extraction module is trained only with perfect cut-out images where sub-objects are arranged in the same order, then the analysis precision is high for such images, but the analysis precision is reduced when handling imperfect cut-out images where the order of sub-objects differs or only part of the object is reflected
Solution Approach 1:
The patent segments the feature value extraction module into multiple local branches, where each local branch is responsible for extracting feature values from specific sub-objects (e.g., head, upper body, lower body). This segmentation allows each branch to specialize in detecting particular sub-objects regardless of their position in the image, thereby improving adaptability to imperfect cut-out images while maintaining high analysis precision for complete images.
2Measurement precision
If the local branches are optimized using indirect review from the analysis module, then the analysis precision is improved, but the regions of sub-objects picked up by each local branch are not precisely determined
Solution Approach 1:
The patent implements a feedback mechanism where the analysis module provides feedback signals to the local branches based on their extraction results. This feedback is used to update the weight parameters of each local branch, enabling them to learn and improve their region-picking accuracy over time. The feedback loop ensures that each local branch receives direct guidance on which regions to focus on, combining both precise region identification and high analysis precision.
Data Source
AI summary
A learning device allocates which feature value of a sub-object is extracted by a module from a group of sub-objects constituting an object of an image to each of modules that extract feature values of the object in the image in a deep neural network that is a learning target. After that, the learning device performs first learning to perform learning of the respective modules so that the respective modules are capable of precisely picking up regions of sub-objects allocated to the modules using information indicating the regions of the sub-objects in an image for each of images and second learning to perform learning of the respective modules so that analysis precision of an image analysis is further improved using a result of the image analysis based on the feature values of the sub-objects picked up by the respective modules.