Dual-Stream Object Detection With Retrieval for Open Categories
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models struggle with accuracy when applied to large-scale image databases outside their training data set, as they are trained on limited data and fail to accurately recognize open data categories.
Innovation Solution
A dual-stream detection approach combining an object detection model and an object retrieval model, where the object detection model performs initial feature extraction and location prediction, and the object retrieval model enhances category prediction using sample images, with confidence levels combined to determine the final category.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If an object detection model is trained on limited data sets, then training efficiency is improved, but detection accuracy on large-scale image databases deteriorates
Solution Approach 1:
The system performs preliminary object detection to locate candidate regions, then conducts detailed retrieval-only analysis on these regions. This preliminary action allows the system to use simpler models for initial screening while reserving computational resources for accurate classification on identified candidates, thereby maintaining both training efficiency and detection accuracy.
Solution Approach 2:
The detection process is segmented into two distinct stages: an object detection stage that identifies candidate regions using location information and confidence levels, and a retrieval stage that performs detailed category classification. This segmentation allows each stage to be optimized independently, with the detection stage focusing on speed and the retrieval stage on accuracy.
2Device complexity
If a single detection model is used, then device complexity is reduced, but the ability to handle open data categories deteriorates
Solution Approach 1:
The system employs a retrieval model that is trained exclusively on retrieval tasks and can handle any category within its training scope. This universal retrieval model works alongside the object detection model to provide accurate classification for open data categories, making the overall system versatile without requiring complex category-specific adaptations.
Solution Approach 2:
The retrieval model acts as an intermediary between the object detection model and the final category classification. It receives candidate regions from the detection model, performs detailed analysis, and returns confidence levels for various categories. This intermediary layer enables the system to handle diverse categories without increasing the complexity of the core detection model.
3Adaptability or versatility
If multi-scale feature extraction is performed, then detection coverage is improved, but computational complexity increases
Solution Approach 1:
The system applies different processing strategies to different scales: multi-scale feature extraction is performed to ensure coverage of objects at various sizes, but the computationally intensive retrieval analysis is applied only to candidate regions identified at relevant scales. This local quality approach ensures detection coverage while managing computational complexity by avoiding unnecessary processing at all scales.
Data Source
AI summary
This application discloses an artificial intelligence-based object detection method and apparatus. The method includes inputting a target image comprising an object to an object detection model; obtaining feature images of different scales from the target image using the object detection model; determining image location information of the object and a first confidence level that the object belongs to each category; acquiring a target region in which the object is located; inputting the target region to an object retrieval model, comparing the target region with sample images of a plurality of categories to obtain a second confidence level that the object belongs to each category; and determining a target category of the object based on the first confidence level and the second confidence level, a sum of the first confidence level and the second confidence level of the target category being the largest of the plurality of categories.