Leveraging robust CNN architectures for real-time object recognition from conveyor belt

Abstract

In the contemporary era, the problem of recognizing undesirable objects and individuals on the conveyor belt is addressed by various architectural or algorithmic approaches. Conveyor belts are an object-carrying medium by which things go in a straight line for transportation. However, sometimes some unwanted objects go through the belt mistakenly, which can be very dangerous. Hence, detection accuracy is much needed to avoid such deadly occurrences and better accuracy can be achieved by performing detection in real time. Furthermore, this upgraded system will assist individuals by lowering the danger of any accidents and provide a real-time example of an airport conveyor belt by detecting any unwanted moving objects with the help of different algorithms and methods of the neural network. Therefore, in this paper, we offer a few algorithms that comprise Convolutional Neural Network (CNN), YOLOv5, YOLOv7, and Vision Transformer (ViT) as well as some Transfer learning methods over a few pre-trained models (VGG16, ResNet50, MobileNetV2) to produce a better strategy on our customized dataset to find out if it can recognize objects and human babies. Our research indicates that our custom dataset closely fits the architecture of YOLOv5, with an accuracy of 88%.