Two-Stage Target Detection Using Salience Area Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current target detection methods face challenges in accurately identifying small targets within large images due to low sensitivity and accuracy, as they fail to utilize inherent characteristics of small targets and often introduce false positives through image interpolation, which increases computational costs and false alarms.
Innovation Solution
A target detection method involving a two-stage classification process: first, image patches are classified using a deep neural network to identify candidate patches, and then salience areas are extracted and feature vectors are constructed to refine the classification using a second model, thereby enhancing sensitivity and accuracy by leveraging inherent target characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image interpolation is used to enhance small targets, then detection coverage is improved, but computational cost and false positives increase
Solution Approach 1:
The image is divided into multiple patches, and only patches containing small targets are selected for further processing. This segmentation approach avoids applying computationally expensive interpolation to the entire image, thereby reducing overall computational cost while maintaining detection coverage for small targets.
Solution Approach 2:
Different processing strategies are applied to different regions: candidate patches undergo interpolation and detailed analysis, while non-candidate patches are processed more simply or skipped. This local quality approach optimizes computational resource allocation, applying high-cost operations only where necessary to detect small targets.
2Productivity
If traditional detection methods are applied to small targets, then processing speed is maintained, but detection sensitivity and accuracy deteriorate
Solution Approach 1:
A first classification model performs preliminary screening to identify candidate patches containing small targets before applying more sensitive but computationally intensive detection methods. This preliminary action filters out non-relevant regions, allowing high-sensitivity processing to be applied only where needed, thus maintaining overall processing speed while improving detection sensitivity.
Solution Approach 2:
The first classification model acts as an intermediary between rapid scanning and detailed small target detection. It provides a filtering mechanism that enables the system to maintain processing speed through fast classification while improving sensitivity through subsequent detailed analysis of only candidate regions.
3Measurement precision
If comprehensive feature analysis is performed on all image patches, then detection accuracy is improved, but processing time increases
Solution Approach 1:
The image processing is segmented into two stages: rapid classification of all patches using a first model, followed by detailed feature analysis only for candidate patches using a second classification model. This segmentation reduces processing time by avoiding comprehensive analysis of all patches while maintaining detection accuracy through thorough analysis of relevant candidates.
Solution Approach 2:
Instead of performing complete feature analysis on all patches (excessive action), the system performs partial analysis on filtered candidate patches. This partial action approach maintains detection accuracy for small targets while significantly reducing overall processing time by limiting detailed analysis to necessary regions only.
Data Source
AI summary
A target detection method a is provided, which relates to the fields of deep learning, computer vision, and artificial intelligence. The method comprises: classifying, by using a first classification model, a plurality of image patches comprised in an input image, to obtain one or more candidate image patches, in the plurality of image patches, that are preliminarily classified as comprising a target; extracting a corresponding salience area for each candidate image patch; constructing a corresponding target feature vector for each candidate image patch based on the corresponding salience area for each candidate image patch; and classifying, by using a second classification model, the target feature vector to determine whether each candidate image patch comprises the target.