Snake Identification System Using Convolutional Neural Networks

Abstract

Computer vision has recently been dominated by Convolutional Neural Networks (CNNs), these are a kind of Artificial Neural Networks (ANNs) mostly employed for image classification and object detection. Identifying a snake species is important when interacting with the species as well as when treating injuries due to envenoming. This task however proves to be a hurdle for the general public. This paper, therefore, sought to solve the problem of misidentification of snake species which often leads to envenoming, and mishandling of snake species by harnessing the power of CNNs together with the portability of mobile devices in developing a mobile application that identifies snake species from images almost in real-time. In implementing this system, the CNN model was trained in Google Collab on a custom-tailored dataset. The images in the dataset were sourced from the internet, and were divided into eight classes which represented eight different snake species. The images were annotated using MakeSense.ai, an online data annotation tool. After annotation the images were piped into the YOLOv5 CNN model on Google Collab for model training. The training process yielded an accuracy of 71% for all the eight classes. After training, the model was converted to a Tensorflow Lite model and exported to Android Studio IDE wherein the rest of the application was developed using Java programming language.