Image Hashing Object Detection With Class-Balanced Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural networks for object detection in image hashing face inaccuracies due to inadequate training on real-world datasets with long-tailed class distributions, leading to skewed neural networks that favor common classes and underperform for rare classes.
Innovation Solution
A unified framework combining representation learning and classifier learning with class-balanced loss generation, using a cumulative learning stage to re-balance the sample distribution by factoring the effective number of samples, thereby improving accuracy for both common and rare classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep neural networks are trained on real-world datasets with long-tailed class distributions, then the network can handle diverse object classes, but the network becomes skewed and favors common classes while underperforming on rare classes
Solution Approach 1:
The training process is segmented into two distinct stages: a representation learning stage that learns general features from all classes, and a classifier learning stage that specifically addresses class imbalance through class-balanced loss. This segmentation allows the network to first capture universal representations and then compensate for long-tail distribution issues without compromising either stage.
Solution Approach 2:
The patent applies parameter changes by introducing class-balanced loss functions that dynamically adjust weighting parameters during training. The loss function incorporates effective number of samples per class, transforming the optimization landscape to penalize misclassifications of rare classes more heavily, thereby changing the network's parameter updates to favor rare class detection.
2Measurement precision
If the neural network is trained to compensate for rare classes, then detection accuracy for rare classes improves, but the network structure becomes more complex and training becomes less efficient
Solution Approach 1:
By separating representation learning from classifier learning, the patent avoids the need for complex compensation algorithms within a single unified network. The representation layer remains simple and universal, while the classifier layer handles class imbalance through loss function adjustments, reducing overall structural complexity.
Solution Approach 2:
The class-balanced loss function provides feedback mechanisms that automatically adjust training based on class distribution. The effective number of samples calculation provides a feedback loop that adapts the loss weights dynamically, eliminating the need for manual compensation algorithms or complex network modifications.
3Productivity
If conventional training methods are used with standard loss functions, then training is simple and fast, but the network produces skewed results favoring common classes
Solution Approach 1:
The patent modifies the loss function parameters by incorporating class-specific weighting based on effective number of samples. This parameter change transforms the standard cross-entropy loss into a class-balanced loss that maintains training efficiency while producing reliable detection across all classes, including rare ones.
Solution Approach 2:
The patent replaces mechanical compensation methods (such as oversampling or complex compensation algorithms) with a mathematical substitution in the loss function. The class-balanced loss function mathematically accounts for class imbalance without requiring data manipulation or complex algorithmic corrections, maintaining training simplicity while improving reliability.
Data Source
AI summary
A method and system of image hashing object detection for image processing are provided. The method comprises the following steps: obtaining image head class input data and image tail class input data differentiated from the head class input data and respectively of two images each of an object to be classified; respectively inputting the head and tail class input data into two separate parallel representation neural networks being trained to respectively generate head and tail features, wherein the representation neural networks share at least some representation weights used to form the head and tail features; inputting the head and tail features into at least one classifier neural network to generate class-related data; generating a class-balanced loss of at least one of the classes of the class-related data comprising factoring an effective number of samples of individual classes; and rebalancing an output sample distribution among the classes at the representation neural networks, classifier neural networks, or both by using the class-balanced loss.