Image Recognition Apparatus Overlapping Candidate Area Discrimination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition techniques often erroneously detect physical bodies or backgrounds similar to human bodies, leading to incorrect detection results, especially when overlapping detection results occur, causing correctly detected human body areas to be deleted as erroneous detections.
Innovation Solution
An image recognizing apparatus comprising a detecting unit, an extracting unit, a classifying unit, and a discriminating unit that identifies candidate areas in an input image, classifies overlapping states, and discriminates whether areas are human body targets based on their likelihoods and positional relations to accurately distinguish between correct and erroneous detections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing image recognition techniques are used to detect human body areas, then detection speed and basic detection capability are improved, but erroneous detections increase when similar shapes (shoulders, legs, backgrounds) are present
Solution Approach 1:
The detection process is segmented into multiple stages: initial detection using HOG features for speed, followed by classification of overlapping detection results, and final discrimination using multiple features including body part relationships. This segmentation allows the system to maintain high detection speed while improving accuracy through progressive filtering of erroneous detections.
Solution Approach 2:
A classification unit is introduced as an intermediary between the initial detection unit and the final discrimination unit. This intermediary classifies overlapping detection results into different types (e.g., single human body, multiple human bodies, human body with background) before final discrimination, enabling more accurate handling of complex detection scenarios and reducing erroneous detections.
2Device complexity
If overlapping detection results are resolved by comparing likelihoods alone, then processing simplicity is maintained, but correct human body areas may be deleted as erroneous detections in cases like children in front of adults or distant persons
Solution Approach 1:
The system performs preliminary classification of overlapping detection results before final discrimination. By pre-classifying the types of overlapping scenarios (single human body, multiple human bodies, human body with background) and establishing appropriate discrimination methods for each type, the system avoids the need for complex real-time decision-making while ensuring correct detections are retained.
Solution Approach 2:
The discrimination process uses multiple parameters beyond simple likelihood comparison, including body part relationships, spatial positions, and classification types of overlapping results. By changing from a single-parameter (likelihood) approach to a multi-parameter approach, the system can distinguish between correct and erroneous detections more accurately without excessive complexity.
3Productivity
If simple likelihood comparison is used for overlapping detection results, then computational efficiency is improved, but detection precision deteriorates when multiple valid human body areas exist
Solution Approach 1:
The discrimination process is segmented into classification-based decision-making and feature-based discrimination. For certain classification types (e.g., single human body), simple likelihood comparison is sufficient and is applied directly. For other types (e.g., multiple human bodies or human body with background), more sophisticated multi-feature discrimination is applied. This segmentation maintains computational efficiency for simple cases while improving precision for complex cases.
Data Source
AI summary
An image recognizing apparatus is equipped with: a detecting unit configured to detect, from an input image, a candidate area for a target of recognition, based on a likelihood of a partial area in the input image; an extracting unit configured to extract, from a plurality of candidate areas detected by the detecting unit, a set of the candidate areas which are in an overlapping relation; a classifying unit configured to classify an overlapping state of the set of the candidate areas; and a discriminating unit configured to discriminate whether or not the respective candidate areas are the target of recognition, based on the overlapping state of the set of the candidate areas and the respective likelihoods of the candidate areas.