Facial Expression Recognition Using Gabor Filter Ensemble
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial expression classification using machine learning technologies faces challenges in achieving high accuracy due to the complexity of facial expressions.
Innovation Solution
A computer-implemented method involving filtering an input image using Gabor filters to generate filter response images, which are then processed by two neural networks (DenseNet) to generate prediction values that are weighted averaged to improve facial expression recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning methods are used for facial expression classification, then the system is simpler to implement, but the recognition accuracy is insufficient due to facial expression complexity
Solution Approach 1:
The patent segments the facial expression recognition task into multiple independent streams: (1) direct classification of the original image, (2) classification of Gabor filter response images at multiple scales and orientations, and (3) classification of Laplacian pyramid images at multiple levels. Each stream is processed by separate neural networks, and their predictions are combined through weighted averaging. This segmentation allows the system to capture different aspects of facial expressions (texture, scale, orientation, frequency) independently, thereby improving overall accuracy while maintaining modular system complexity.
Solution Approach 2:
The patent employs a composite approach by combining multiple types of image representations (original images, Gabor filter responses, Laplacian pyramid decompositions) and multiple neural network architectures (DenseNet variants) into a unified ensemble system. Each component contributes different features and perspectives to the final classification decision, similar to how composite materials combine different substances to achieve superior properties. The weighted averaging mechanism integrates these diverse predictions to produce a more accurate and robust classification result.
2Measurement precision
If multiple neural networks and filter processing are used to improve accuracy, then recognition precision increases, but processing time and computational resources increase
Solution Approach 1:
The patent performs preliminary processing by generating Gabor filter response images and Laplacian pyramid decompositions before the main classification task. These preprocessed images capture essential features (texture, edges, frequency components) at multiple scales and orientations, which are then fed into the neural networks. This preliminary action reduces the computational burden during inference by providing ready-to-use feature representations, thereby balancing accuracy improvement with processing efficiency.
Solution Approach 2:
The patent implements a multi-scale and multi-orientation filtering approach that processes images at more detail levels than a single-scale system would require. By using Gabor filters at multiple scales and orientations, and Laplacian pyramid at multiple levels, the system captures comprehensive facial expression features. The weighted averaging mechanism selectively combines these partial results, using more processing resources only when they contribute to improved accuracy, thus achieving a balance between thoroughness and efficiency.
Data Source
AI summary
A computer-implemented method of recognizing a facial expression of a subject in an input image is provided. The method includes filtering the input image to generate a plurality of filter response images; inputting the input image into a first neural network; processing the input image using the first neural network to generate a first prediction value; inputting the plurality of filter response images into a second neural network; processing the plurality of filter response images using the second neural network to generate a second prediction value; weighted averaging the first prediction value and the second prediction value to generate a weighted average prediction value; and generating an image classification result based on the weighted average prediction value.