Multiple Object Recognition from Smart Document Images using YOLOv5s

Abstract

Multiple object recognition from various ID proofs will give more security as well as authentication of the data. In general, various ID proofs showcase objects in various ways, so manually identifying common objects is a time-consuming process. Some of the existing work like YOLOv3, YOLOv5, CNN, and Faster RCNN, all concentrate on one single object detection from the same dataset category, while the proposed work concentrates on heterogeneous datasets with the YOLOv5s model. The proposed work will automate the multiple object detection and recognition from various government id proofs as well as overcoming overlapped object recognition. The proposed model includes deep YOLOv5, which contains 19 convolution layers and 5 pooling layers. The proposed model detects the object from 1000 manually collected various government ID proofs like driving licence, PAN card, Aadhaar card, and voters' ID. The model is trained with 750 datasets and tested with 250 datasets, and finally validated with 50 datasets. The model clearly detects and recognize the name, unique identification number, date of birth, and photograph with 94.6% accuracy. Also, the model recognizes overlapping signatures with better accuracy.