Facial Expression Recognition Using Hybrid ILSGA-BP Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial expression recognition technologies face challenges in accurately classifying emotions under varying expressions, lighting conditions, and occlusions, with existing methods often resulting in low accuracy and misclassification, especially for emotions like disgust and surprise.
Innovation Solution
The method employs a hybrid approach combining genetic algorithms (GA) and iterated local search (ILS) with back-propagation (BP) to optimize neural network weights for facial expression recognition, using radial curves and Cubic Bezier curves to extract features from the eyes and mouth, and classifies expressions into seven emotion categories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional back-propagation algorithm is used for training neural network, then training process is simple, but recognition accuracy is low especially for emotions like disgust and surprise
Solution Approach 1:
The patent combines genetic algorithm (GA) and iterated local search (ILS) with back-propagation (BP) to create a hybrid ILSGA-BP algorithm. The GA provides global search capability to avoid local optima, ILS refines solutions through local search, and BP performs gradient-based optimization. This merging of multiple optimization approaches resolves the contradiction by achieving high recognition accuracy through comprehensive search while maintaining structured training process.
Solution Approach 2:
The training process is segmented into distinct phases: GA phase for global optimization of network weights, ILS phase for local refinement, and BP phase for fine-tuning. Each segment addresses specific aspects of the optimization problem, with GA handling broad search, ILS improving local solutions, and BP performing precise adjustments. This segmentation allows the system to achieve high accuracy without overwhelming complexity at any single stage.
2Measurement precision
If more features are extracted from facial parts to improve recognition accuracy, then computational complexity increases
Solution Approach 1:
The patent extracts specific geometric features from key facial parts (eyes and mouth) using radial curves and Cubic Bezier curves. Instead of processing entire facial images or all possible features, the method selectively extracts curvature and shape characteristics from regions most indicative of emotional expression. This extraction approach achieves high recognition accuracy while reducing computational complexity by focusing only on the most discriminative features.
Solution Approach 2:
The patent applies different feature extraction methods to different facial parts based on their local characteristics. Radial curves are used to capture the shape of eyes, while Cubic Bezier curves model the mouth contour. Each facial region is processed with the most appropriate mathematical representation, optimizing the balance between feature quality and computational efficiency for that specific local area.
Data Source
AI summary
The disclosure provides an automatic geometric method for analyzing and recognizing human facial expressions based on extracted features using a neural network. The extracted features are obtained from extracted right eye, left eye and mouth using radial curves and Cubic Bezier curves. Facial expressions are classified into seven different emotion categories, namely surprise, happiness, disgust, neutral, fear, sadness and anger. The method employs an algorithm which hybridizes genetic algorithms (GA) and iterated local search (ILS) with back-propagation algorithm (BP). The BP is used to train the neural network. The GA and ILS are used to enhance and optimize initial weights of the BP.