Image Classification System for Zero-Shot Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In generalized zero-shot recognition, the model overfits to seen labels used in training data, leading to less accurate recognition of unseen labels in test data, as labels of unseen images are not used in the learning process.
Innovation Solution
An image classification system that computes a known-image probability and likelihoods for both seen and unseen images, with a correction mechanism using the known-image probability to improve label estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the model is trained using only seen labels from training data, then the model can learn from available data, but the model overfits to seen labels and fails to accurately recognize unseen labels in test data
Solution Approach 1:
The patent introduces a new dimension to the problem by computing known-image probabilities separately from the standard likelihood computation. Instead of only computing likelihoods in the label space, the system adds a probability dimension that measures how likely an input image is to be a known image versus an unseen image. This dimensional expansion allows the model to distinguish between seen and unseen cases, resolving the overfitting problem without increasing model structural complexity.
Solution Approach 2:
The patent modifies the recognition parameters by introducing known-image probabilities as an additional parameter. The final recognition result is obtained by combining the computed likelihood with the known-image probability through multiplication. This parameter change allows the system to adjust its confidence in unseen label recognition based on the probability that the input is actually a known image, thereby improving reliability without complicating the model architecture.
2Adaptability or versatility
If the model computes likelihoods for both seen and unseen images, then comprehensive classification is achieved, but the recognition accuracy for unseen labels decreases due to overfitting to seen labels
Solution Approach 1:
The patent implements a feedback mechanism where the known-image probability is computed based on the input image features and then used to correct the likelihood computation. The system feeds back the probability information to adjust the final recognition result, creating a self-correcting mechanism that improves measurement precision while maintaining adaptability across seen and unseen images.
Solution Approach 2:
The known-image probability acts as an intermediary between the likelihood computation and the final recognition result. Instead of directly outputting the likelihood as the recognition result, the system introduces this probability intermediary that mediates the relationship between the input image and the predicted label, thereby improving accuracy by accounting for the possibility that the input is a known image.
Data Source
AI summary
An image classification system 10 includes: a probability computation means 11 which computes a known-image probability, which is the probability that an input image corresponds to a known image associated with a seen label that indicates the class into which content indicated by the known image is classified; a likelihood computation means 12 which computes both the likelihood that content indicated by the input image is classified into the same class as content indicated by an unseen image associated with an unseen label, and the likelihood that the content indicated by the input image is classified into the same class as the content indicated by the known image; and a correction means 13 which corrects each computed likelihood using the computed known-image probability.