Object Evaluation Reweighting for Imbalanced Level Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object evaluation methods, particularly those using computer models, suffer from inaccurate and unreliable assessments due to unbalanced training data distribution, where higher-level objects have abundant data while lower-level objects lack sufficient representation, leading to poor evaluation accuracy.
Innovation Solution
An object evaluation method that adjusts weightings based on the distribution of known object levels in the training data to refine the confidence ratings, using neural network classifiers and coordinate transformations to enhance accuracy and reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If computer models are used for automatic object evaluation, then evaluation efficiency is improved, but evaluation accuracy deteriorates due to unbalanced training data distribution
Solution Approach 1:
The patent changes the parameter of training data distribution by introducing reweighting mechanisms. It calculates importance weights for different object levels based on their frequency distribution, then adjusts the loss function and sampling strategies to balance the contribution of underrepresented levels, thereby improving evaluation accuracy while maintaining automated efficiency
Solution Approach 2:
The patent applies partial action by focusing computational resources on the minority class samples that need improvement. Instead of uniformly processing all data, it implements oversampling for underrepresented object levels and applies higher weighting to their loss contributions, addressing the specific weakness in the training data distribution without reprocessing the entire dataset
2Measurement precision
If more training data for lower-level objects is collected, then evaluation accuracy is improved, but data collection cost and time increase
Solution Approach 1:
The patent performs preliminary action by proactively identifying the distribution imbalance problem during the training phase and applying corrective measures before final evaluation. It calculates the distribution statistics of training data, determines reweighting factors in advance, and integrates these into the training process, avoiding the need for time-consuming additional data collection later
Solution Approach 2:
The patent introduces an intermediary mechanism in the form of a reweighting system that mediates between the existing unbalanced training data and the model training process. This intermediary layer adjusts the effective contribution of different samples through calculated weights, bridging the gap without requiring physical addition of more training samples
Data Source
AI summary
An object evaluation method and an object evaluation device. The object evaluation method includes: obtaining an image of a first object to be evaluated; determining first level information by means of a first object level classifier according to the image of the first object to be evaluated; determining adjustment weights of various object levels according to distribution proportions of the object images of the various object levels in a plurality of object images having the same object type as that of the image of first object to be evaluated and having known object levels, wherein an object level having a relatively high distribution proportion has a relatively low adjustment weight; and respectively generating second rating confidences of the various object levels according to first rating confidences and the adjustment weights of the various object levels to form second level information.