Image Database Search With Iterative Feedback for Edge-Case Curation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of identifying suitable images for training neural networks, particularly in edge cases, from large databases is exacerbated by the difficulty in finding relevant images amidst billions of available images, and existing methods struggle to efficiently curate diverse and varied image datasets for robust training.
Innovation Solution
A system and method for searching an image database using a user interface that receives user inputs, either in text or image form, to refine image selections through iterative user feedback, leveraging a CLIP model to identify and retrieve images that match or differ from initial search results, allowing for the creation of a refined dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large image database is searched manually to find edge case images for training, then the diversity and robustness of the training dataset can be improved, but the time and effort required to curate the dataset increases significantly
Solution Approach 1:
The patent introduces an intermediary system comprising a trained neural network and a user interface that mediates between the large image database and the user's search needs. The system automatically processes the database using the trained network to identify edge cases, while the user interface allows users to specify search criteria and review results without manually examining each image. This intermediary automation resolves the contradiction by reducing curation time while maintaining dataset robustness.
Solution Approach 2:
The system enables self-service dataset curation by allowing users to independently search and select edge case images through the user interface without requiring manual image-by-image analysis. The trained neural network automatically screens the database according to user-specified criteria, empowering users to curate robust datasets efficiently on their own without extensive manual effort.
2Productivity
If synthetic image generation is used to create training data, then the scalability of data generation is improved, but the realism and variety of the generated images decrease
Solution Approach 1:
Instead of generating synthetic images from scratch, the system copies and selects from real images in the database that match desired edge case characteristics. The trained neural network identifies and retrieves actual images containing rare or difficult scenarios, preserving their inherent realism while providing scalable access to diverse training examples without the quality degradation associated with synthetic generation.
3Ease of operation
If traditional image search methods are used to find edge cases, then the simplicity of the search process is maintained, but the precision of identifying relevant edge case images decreases
Solution Approach 1:
The system performs preliminary action by pre-training the neural network on the image database before the actual search occurs. This pre-processing step enables the network to understand and identify edge case characteristics automatically during the search phase. Users can then perform simple searches without sacrificing precision, as the heavy lifting of accurate identification has already been prepared in advance through network training.
Solution Approach 2:
The patent replaces manual mechanical image analysis with an automated neural network-based system. Instead of users manually examining images to identify edge cases, the trained network automatically detects and ranks relevant images based on user criteria. This substitution maintains ease of operation for users while dramatically improving the precision of edge case identification through intelligent automation.
Data Source
AI summary
The present disclosure is directed to searching an image database. A system may include at least one processor comprising circuitry and a memory, wherein the memory includes instructions that, when executed by the circuitry cause the at least one processor to: receive a first user input including one or more scenario descriptors, identify in a database an initial plurality of images based on the first user input and display the initial plurality of images to a user. The processor may be further configured to receive a second user input wherein the second user input identifies at least one of the initial plurality of images. The at least one processor may identify a refined plurality of images in the database based on the one or more scenario descriptors in combination with the second user input and display the refined plurality of images to the user.