Image-Based Hand Tools and Accessories Recognition by ResNet50

Abstract

The objective of this research is to create a computer system which can recognize various kinds of hand-tools by using only a single image.  The developed system is called “Hand-tool and accessory image recognition system or (HTAIRS)”.  The system consists of 4 main modules, namely: 1) dataset training, 2) image acquisition, 3) image recognition, and 4) result presentation modules.  The system employs the convolutional neural networks (CNN) called “ResNet50”, which is a toolbox in MATLAB software.  The developed system creates its own dataset called “Hand Tools Dataset”, which consists of 165 different video clips in 110 hand tool categories and 600 images each. The HTAIRS separates 500 images for training dataset and 100 images for evaluating the system.  The accuracy of the training system is 99.30% and the accuracy that of the evaluating is also 99.30%.  The system’s average access time are is 0.8549 Seconds per image.