Enterprise AI Analysis

EnhancedKinshipNet utilizes a pre-trained ResNet50 for high-level visual feature extraction from parent and child images, producing 2048-dimensional feature vectors. A lightweight transformer-based module integrates these features, capturing relational relationships and reducing dimensionality to 512, with learnable positional embeddings to distinguish inputs. This module uses 8 attention heads to improve feature interactions, followed by mean pooling and an unprojection layer to restore 2048-dimensional vectors. Finally, a shallow Multi-Layer Perceptron (MLP) classifier with dropout determines the kinship category, providing a binary prediction of familial relation. The model employs 10-fold cross-validation, Adam optimizer, a batch size of 32, and a learning rate of 1x10^-4, with various data augmentation techniques to ensure robust performance.

The EnhancedKinshipNet model was rigorously evaluated on the KinFaceW-I and KinFaceW-II benchmark datasets, both individually and combined. On KinFaceW-I, the model achieved an accuracy of 94.44%. For KinFaceW-II, a higher accuracy of 96% was attained, reflecting the dataset's greater consistency in image conditions. When both datasets were combined, the overall accuracy was 93.46%. These results demonstrate competitive performance against state-of-the-art methods across all four kinship relationship classes (Mother-Daughter, Father-Daughter, Mother-Son, Father-Son), validating the model's effectiveness in accurately classifying familial relationships despite age and gender variations. The architecture's ability to optimize feature extraction and fusion contributed significantly to these high accuracy scores.

To address the 'black box' nature of deep learning, EnhancedKinshipNet integrates Explainable AI (XAI) techniques, specifically Grad-CAM and LIME. Grad-CAM generates heatmaps that highlight the facial regions most influential in the model's predictions, consistently focusing on eyes, nose, and mouth—areas crucial for human kinship identification. This visual evidence enhances trust by showing that predictions are based on semantically meaningful features. LIME further offers instance-level interpretability by locally approximating predictions with linear models, using superpixel-based heatmaps to show positive and negative contributions of image regions. These XAI insights validate the model's reasoning, particularly for same-gender relationships, and identify areas for improvement in cross-gender and lower-resolution scenarios, reinforcing the model's credibility and applicability in sensitive, human-centric domains.