Overview

This project focuses on human emotion detection using ViT and CNN models with quantization for efficient AI. The main components are:

• References:
  • My-Model
  • Model Visualization, Patch Encoder, and Transfer Learning
  • Hugging Face Model with WandB
  • Cutmix Augmentation and TensorFlow Record
  • ONNX Quantization and TensorFlow Lite Model Accuracy Evaluation and Conversion
  • Human Face Recognition
  • API
  • Data


• Visualizing Model Features, GradCAM, and Patch Encoding:
  • Use VGG16 to visualize feature maps.
  • Implement GradCAM with EfficientNetB5.
  • Extract and visualize image patches using TensorFlow.
  • Define custom ResNet34 for deep feature extraction.


• Hugging Face Vision Transformer (ViT):
  • Initialize ViT model configuration.
  • Load and fine-tune pre-trained ViT model from Hugging Face.
  • Define custom architecture with additional layers for classification.
  • Track experiments with WandB.
  • Train ViT-based model.


• Data Augmentation and TensorFlow Records:
  • Define augmentation layers: rotation, flip, contrast.
  • Apply CutMix augmentation.
  • Create TensorFlow Records for efficient data storage.


• FastAPI Deployment and ONNX Quantization:
  • Deploy emotion detection model using FastAPI.
  • Optimize model with ONNX quantization.


• Face Recognition Integration:
  • Add context to emotion detection with face recognition.
  • Enhance user engagement and security.
  • Combine emotion detection and face recognition for better analysis.
  • Generate detailed reports with face recognition data.
  • Unified API for emotion detection and face recognition.

We appreciate the support from the open-source community, particularly contributions to OpenCV.

This Project’s GitHub Repository