Bounded ReLU Image Feature Extraction to Reduce Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning (DL) based image feature extraction methods face challenges such as limitations in processing grayscale and low-resolution images, overfitting, and biased feature extraction due to considering only positive or all negative features, leading to inaccurate results.
Innovation Solution
A method and system utilizing a bounded Rectified Linear activation Unit (ReLU) based Bounded-Rectifier Network (B-RectNet) for image feature extraction, which includes preprocessing to remove anomalies and filters negative features using lower and upper negative bounds, reducing overfitting and enhancing feature extraction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional DL models consider all negative features along with positive features, then feature extraction is performed, but bias in feature extraction occurs leading to inaccurate results
Solution Approach 1:
The patent extracts and separates negative features from positive features using the bounded ReLU activation function. By introducing lower and upper negative bounds, the model selectively extracts useful negative features while filtering out harmful ones, eliminating the bias that occurs when all negative features are treated uniformly with positive features.
Solution Approach 2:
The patent changes the parameter range of the ReLU activation function from the conventional [0, ∞) to a bounded range with specific lower and upper negative bounds. This parameter modification allows the activation function to handle negative features differently, enabling selective extraction based on bounded ranges and improving feature extraction accuracy while reducing bias.
2Device complexity
If conventional ReLU activation function is used, then processing is simplified, but overfitting occurs due to unlimited negative values
Solution Approach 1:
The patent modifies the ReLU activation function by introducing bounded parameters (lower and upper negative bounds) to constrain the output range. This parameter change prevents the activation function from producing unlimited negative values that cause overfitting, while maintaining relative computational simplicity compared to unbounded ReLU.
Solution Approach 2:
The patent applies bounded constraints to the ReLU activation function in advance to prevent overfitting before it occurs during training. By预先 setting lower and upper negative bounds, the model is cushioned against the harmful effects of unlimited negative values that would otherwise lead to overfitting.
3Ease of operation
If conventional DL models process images directly, then processing is straightforward, but limitations occur in processing high-resolution images
Solution Approach 1:
The patent segments the image processing task into multiple stages: preprocessing to remove anomalies, bounded ReLU activation to filter features, and hierarchical feature extraction. This segmentation allows the model to handle high-resolution images effectively by breaking down the complex processing into manageable steps that preserve important features while reducing computational burden.
Solution Approach 2:
The patent performs preliminary actions including image preprocessing to remove anomalies and apply bounded ReLU activation before main feature extraction. This preliminary processing prepares high-resolution images by filtering out noise and irrelevant features, making subsequent processing more effective and enabling the model to handle high-resolution inputs without straightforward processing limitations.
Data Source
AI summary
The present disclosure provides a model for deep learning based image feature extraction considering a range of useful negative features. Conventional methods are either considering only positive features or considering all negative features along with positive features which leads to bias in feature extraction. The present disclosure overcomes the problem of the conventional methods using a bounded Rectified Linear activation Unit (B-ReLU) activation function based Bounded-Rectifier Network (B-RectNet). Initially, the present disclosure receives an image pertaining to an object. Further, the received image is preprocessed to remove a plurality of anomalies associated with the image a preprocessing technique. Further, a plurality of image features are extracted based on the preprocessed image using a trained B-RectNet. The bounded ReLU activation function filters a plurality of negative features based on a lower negative bound value and an upper negative bound value before inputting a plurality of feature values to a subsequent layer.