Classification of Weapons using Convolution Neural Networks Suitable for Portable Devices

Abstract

This article aims to study the process of developing a process for classifying 8 types of weapons from image data using deep learning models, which contribute to the process of using computer vision technology to recognize images and describe weapons accurately and efficiently automatically. This process eliminates errors in analyzing large amounts of image data with human error and reduces the data size for use on portable devices. In this article, there are also models developed to distinguish specific subclasses of aircraft that can be equipped with combat equipment and general aircraft. These aircraft are difficult to classify due to similar physical characteristics, such as color, size, and shape. Therefore, this paper uses TensorFlow and Keras, a library for developing machine learning with artificial convolutional neural network algorithms. In this work, the convolutional neural network is used for the model. The performance of the deep learning model is also improved along with a comparison of 8 pre-trained models. According to the results of this study, the EfficientNet-B0 model has a weapon classification accuracy from image data of 94.01%. Additionally, this work uses features of the TensorFlow Lite library to convert parameters into models. The model is processed using integer-based computation to use quantization principles to reduce the size of the parameters to compare the features, size, latency, and accuracy with parameters The final model is mobile-compatible and is suitable for use on portable devices. The experimental results of the int8 model are the smallest, with an accuracy of 88.64%.