Hybrid Neural Network Architecture for Multi-Label Object Recognition using Feature Fusion
Divya Meena Sundaram, Veeramachaneni Gayathri siva sameeraja, Nagineni Sai Lasya, Meda Sathvika, Veluru Harshitha, J Sheela · Procedia Computer Science · 2022
Multi-label image recognition has gained significance in recent years due to its increase in employment in numerous domains. In multi-label object recognition, objects are recognized by creating a bounding box across every object present in the image and classifying the objects into specific labels. In this project, we propose a hybrid method implementing a transfer learning-based approach using four different Convolutional Neural Network (CNN) models namely Single-shot Detector (SSD), You Only Look Once (YOLO), RetinaNet, and, Faster R-CNN. Further, feature fusion is applied to learn all the image features from different layers. Since the transfer learning approach is utilized, we can train the pre-trained model precisely even with small amounts of data. Through this hybrid approach, we can achieve better accuracy in object recognition in the case of multi-labeled images compared to traditional methods