Image Retrieval Accuracy via Local Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image retrieval technologies face challenges in accurately distinguishing between candidate images due to similar global features caused by matching background environments and target object clothing, leading to reduced retrieval accuracy.
Innovation Solution
The method involves performing face and limb detection on video frames to extract face and limb features, which are then used to create reference motion trajectories, enhancing the accuracy of image retrieval by focusing on specific object features rather than global information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If global image features are used for image retrieval, then the method is simple to implement, but the retrieval accuracy deteriorates when background environments or clothing are similar
Solution Approach 1:
The patent segments the image into multiple body parts (face, upper body, lower body, limbs) and extracts features for each part separately. This segmentation allows the system to focus on discriminative local features rather than global features, resolving the contradiction by maintaining implementation simplicity through modular feature extraction while improving accuracy through part-specific feature comparison.
Solution Approach 2:
The patent extracts specific local features (face features, limb features, body part features) from the global image. By taking out and focusing on these key local features that are less likely to be similar across different images, the system achieves better retrieval accuracy while maintaining a relatively simple implementation through targeted feature extraction.
2Productivity
If only global image features are extracted, then the processing speed is fast, but the ability to distinguish between similar images deteriorates
Solution Approach 1:
The patent divides the image processing task into multiple parallel segmentation operations (face detection, limb detection, body part detection). This segmentation enables the system to process multiple features simultaneously rather than sequentially, maintaining processing speed while improving distinction capability through multi-dimensional feature comparison.
Solution Approach 2:
The patent merges multiple local features (face features, limb features, body part features) into a comprehensive feature representation. By combining these features in a unified retrieval framework, the system achieves both fast processing through efficient feature fusion and high reliability through multiple distinguishing feature dimensions.
3Measurement precision
If multiple local features are extracted and compared, then the retrieval accuracy improves, but the system complexity increases
Solution Approach 1:
The patent segments the feature extraction process into independent modules for different body parts. Each module (face feature extraction, limb feature extraction, body part feature extraction) operates independently and can be implemented using standard computer vision techniques, reducing system complexity while enabling multiple local features to be extracted and compared for high retrieval accuracy.
4Use of energy by moving object
If global features are used, then the computational resources required are low, but the feature discriminability is insufficient
Solution Approach 1:
The patent extracts only the most discriminative local features (face, limbs, body parts) rather than processing the entire global image. This selective extraction reduces computational resource consumption by focusing on key regions while maintaining high feature discriminability through the unique characteristics of each body part.
Data Source
AI summary
This application provides a data processing method performed by an electronic device to improve accuracy of image retrieval. The method includes: respectively performing face detection on at least one reference video frame included in a reference video based on a specified target object type, to obtain at least one face image; extracting a face feature from the face image; respectively performing limb detection on the reference video frame based on the target object type, to obtain at least one limb image; extracting a limb feature from the limb image; and obtaining at least one reference motion trajectory based on each face image and each limb image, each reference motion trajectory comprising a plurality of limb features and a plurality of face features of a target object.