Dynamic Patch Size Adjustment for Image Search Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques that compare feature values acquired by collecting local feature values by patch unit face challenges in maintaining search accuracy when the size of the subject in query and reference images differ, leading to decreased similarity scores and reduced search efficiency.
Innovation Solution
An image processing apparatus and method that dynamically adjusts patch sizes based on the size of the object region in both query and reference images, ensuring that the proportion of the subject in each patch remains consistent across images, thereby enhancing search accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a patch having a predetermined size is uniformly set for the query image and the reference image, then the processing load is reduced and search speed is improved, but the degree of similarity between feature values decreases when the subject size differs between images
Solution Approach 1:
The patch size is dynamically adjusted based on the detected size of the subject in each image. Instead of using a fixed predetermined patch size, the system determines patch size adaptively according to the actual subject dimensions, ensuring that each patch contains a consistent proportion of the subject regardless of the overall image or subject size variations.
Solution Approach 2:
The patch size parameter is changed according to the detected subject size. The system modifies the patch size parameter dynamically to maintain consistent subject proportion within patches, thereby preserving feature value similarity accuracy while still achieving efficient patch-based processing.
2Measurement precision
If local feature values are compared in a pixel unit, then the search accuracy is improved, but the processing load on the computer increases and search speed becomes slower
Solution Approach 1:
The image is segmented into multiple patches instead of processing pixel by pixel. By dividing the image into manageable patch units and computing feature values for each patch, the system reduces the total number of comparisons needed while maintaining meaningful local feature representation, thus improving search speed without completely sacrificing accuracy.
Solution Approach 2:
The system transitions from pixel-level (1D array of pixels) to patch-level (2D grid of patches) feature comparison. This dimensional change allows aggregation of multiple pixel features into a single patch feature vector, reducing computational complexity while preserving spatial relationships and local characteristics.
Data Source
AI summary
To reduce a decrease in search accuracy in a technique for comparing feature values acquired by collecting a local feature value by patch unit, the present invention provides an image processing apparatus 10 including: an image processing unit 11 that detects an object region occupied by an object in an image; a patch size decision unit 12 that decides a patch size, based on a size of the object region; a patch unit feature value-computation unit 13 that sets a plurality of patches having the patch size described above in the object region, and computes a patch unit feature value acquired by collecting a local feature value in a patch for each patch; and a search unit 14 that searches for an image similar to a query image from among a plurality of reference images by using the patch unit feature value.