Multi-task Image Classifier for Ordered Value Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks are inefficient in classifying images with continuous characteristics, as they rely on single thresholds and labels, leading to misclassification and suboptimal training due to imbalanced data sets and inability to represent continuous characteristics effectively.
Innovation Solution
A multi-task image classifier is employed, which uses multiple labels and thresholds to derive ground truth labels representing ranges of values, allowing for more accurate classification and robust training by analyzing images based on multiple predictions rather than single outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single threshold and single label are used for image classification, then the classification process is simple, but the accuracy is reduced due to inability to represent continuous characteristics effectively
Solution Approach 1:
The patent divides the continuous characteristic range into multiple segments by establishing multiple thresholds (e.g., first threshold, second threshold, third threshold) that partition the ordered values into distinct ranges. Each threshold creates a binary classification task, transforming a single complex classification problem into multiple simpler binary problems that collectively capture the continuous characteristic more accurately.
Solution Approach 2:
The patent transitions from a single-dimension classification (one threshold, one label) to a multi-dimensional classification framework by introducing multiple thresholds and corresponding multiple labels. This dimensional expansion allows the system to represent continuous characteristics through multiple binary dimensions, improving measurement precision while maintaining manageable complexity.
2Measurement precision
If multiple labels and thresholds are used for image classification, then the accuracy is improved, but the training data becomes imbalanced and more complex to process
Solution Approach 1:
The patent performs preliminary derivation of multiple labels from each single tagged value before training begins. By pre-computing the binary classification results for each threshold based on the ground truth ordered value, the system creates balanced training datasets where each binary classifier receives appropriately labeled examples, preventing data imbalance issues before training starts.
Solution Approach 2:
The patent creates a multi-functional classification system where a single image with one ordered value generates multiple training examples across different binary classification tasks. Each threshold-based label serves a specific function in capturing different aspects of the continuous characteristic, making the training process more robust despite increased complexity.
3Productivity
If a single label is used for training, then the training process is efficient, but mislabeled data has a greater impact on classification performance
Solution Approach 1:
The patent segments the single classification task into multiple binary classification tasks, each with its own label derived from a specific threshold. This segmentation distributes the impact of potential mislabeling across multiple independent tasks, so that errors in one threshold-based label do not catastrophically affect the entire classification system.
Solution Approach 2:
The patent changes the labeling parameter from a single ordinal value to multiple binary indicators based on threshold comparisons. This parameter transformation allows the system to maintain training efficiency while improving reliability, as the binary labels provide more granular and distributed information about the continuous characteristic.
Data Source
AI summary
In an approach for classifying an image containing a continuous characteristic using an image classifier, an image tagged with a single value is received, wherein the single value is from a set of ordered value. A processor sets at least two thresholds based on the set of ordered values. A processor derives at least two labels according to the single value tagged in the image and the at least two thresholds. A processor runs image through an image classifier. A processor receives at least two predictions from the image classifier based on the at least two labels. A processor adjusts the image classifier based on differences between the at least two predictions and the at least two labels.