Spatial Feature Map Image Similarity for Cropped and Edited Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional digital image similarity techniques are sensitive to differences such as cropping, localized edits, resizing, and format changes, leading to failures in identifying visually similar images, especially in scenarios involving low-level processing artifacts.
Innovation Solution
A visual similarity system employing machine-learning models, specifically convolutional neural networks, to generate spatial feature maps and similarity scores by comparing images at multiple intermediate neural network layers, addressing subtle differences caused by resizing, compression, and format changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional digital image similarity techniques are used, then the system can process images quickly with simple algorithms, but the system fails to identify visually similar images when differences exist due to cropping, resizing, compression, or format changes
Solution Approach 1:
The patent transforms images into a feature space using neural network encoders, changing the parameter representation from raw pixel values to learned feature vectors. This transformation makes the similarity comparison robust to processing artifacts like compression and resizing, as the feature space captures semantic content rather than low-level variations
Solution Approach 2:
The patent introduces spatial feature maps as an intermediary representation between the input images and the similarity score. These feature maps capture local spatial relationships and are used to compute a more robust similarity metric that accounts for localized edits and cropping while maintaining accuracy for visually similar images
2Reliability
If conventional similarity techniques are used, then the system maintains simplicity in implementation, but the system produces incorrect similarity assessments when images undergo low-level processing artifacts
Solution Approach 1:
The patent divides the image comparison process into multiple stages: encoding into feature vectors, generating spatial feature maps, and computing similarity scores. This segmentation allows each component to specialize in a specific aspect of similarity assessment, improving overall reliability while managing complexity through modular processing
Solution Approach 2:
The patent moves the comparison from the traditional pixel-space dimension to a learned feature-space dimension using neural networks. This dimensional transformation enables the system to capture semantic similarities that are invariant to processing artifacts, significantly improving reliability despite the increased computational complexity
3Reliability
If the system uses machine-learning models with spatial feature maps, then the system accurately identifies visually similar images despite processing artifacts, but the computational complexity and processing time increase
Solution Approach 1:
The patent performs preliminary encoding of images into feature vectors and generation of spatial feature maps before the actual similarity comparison. This preliminary processing organizes the data in a way that facilitates efficient comparison operations, reducing the computational burden during the similarity assessment phase
Data Source
AI summary
Digital image visual similarity determination techniques are described. In implementations, a search result is generated based on visual similarity of a plurality of digital images with respect to an input digital image. The search result is generated by locating a plurality of candidate digital images from the plurality of digital images based on a search, calculating spatial feature maps for the input digital image and the plurality of candidate digital images using respective layers of one or more neural networks, and forming a plurality of similarity scores by comparing the spatial feature maps from the plurality of candidate digital images, respectively, with the spatial feature maps for the input digital image.