Topological Order Distance Image Retrieval Scale Rotation Invariance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image retrieval methods struggle with efficiently retrieving images affected by scale, rotation, and quality variations, particularly when image resolution is low or irregular, and fail to effectively capture topological relationships between features.
Innovation Solution
The method involves extracting key features, determining their topological relationships, and using a topological order distance (TOD) to identify images in a database, which includes generating sequences of distances and angles in polar coordinates and comparing these to find matching images efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If SIFT algorithm is used for feature extraction, then rotation and scale invariance is achieved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent segments the feature extraction process into two stages: first using a fast detector (such as interest point detector or blob detector) to identify candidate keypoints, then applying SIFT only at these segmented locations. This segmentation allows the system to achieve SIFT's rotation and scale invariance properties while avoiding the computational burden of applying SIFT throughout the entire image processing pipeline.
2Productivity
If SURF algorithm is used for feature extraction, then processing speed is improved, but performance degrades when image resolution is low
Solution Approach 1:
The patent dynamically adjusts processing parameters based on image resolution. For low-resolution images, it modifies the detection parameters and feature extraction settings to optimize performance. This may involve adjusting the number of octaves, the scale space sampling, or the matching criteria to ensure reliable detection and matching even when image details are limited.
3Device complexity
If vocabulary tree scheme is used for image retrieval, then hierarchical quantization is achieved, but noise level increases and quantization errors are embedded
Solution Approach 1:
The patent applies hierarchical quantization selectively rather than to all features. It uses the vocabulary tree approach for certain feature dimensions while maintaining higher precision for others, or applies it only after initial filtering. This partial application reduces the accumulation of quantization errors while still benefiting from the hierarchical organization and computational efficiency where applicable.
4Adaptability or versatility
If conventional CBIR methods are used for image retrieval, then color, texture, and shape comparison is achieved, but limitation occurs when handling images with scale or rotation variations
Solution Approach 1:
The patent performs preliminary normalization and preprocessing of images before applying conventional CBIR methods. This includes detecting and correcting for scale and rotation variations in advance, or transforming images into a common reference frame. By addressing scale and rotation issues beforehand, the subsequent color, texture, and shape comparisons can proceed with higher reliability, combining the strengths of both conventional CBIR and transformation-invariant approaches.
Data Source
AI summary
An efficient image retrieval method utilizing topological order distance (TOD) algorithm is disclosed. In one aspect, an image recognition method may include steps of receiving one or more input images; extracting one or more key features on the input images; determining an origin and one or more feature points on the key features on the input images; establishing a topological relationship in each dimension between the feature points on the key features; and utilizing the topological relationship in each dimension between the feature points to identify the input images in an image database. Comparing with the conventional image retrieval techniques, the TOD algorithm is advantageous because it is more efficient, scale and rotation invariant, and tolerant to affine and perspective transformation.