IoT based animal classification system using convolutional neural network

Abstract

The kingdom “Animalia” is used to represent all living creatures on the planet earth, which is fallen into six categories. The language is the most common factor to divide humans and animals. Numerous classification techniques can be used for classification purposes, and the classification commonly can be done acoustically and visually. The classification systems are playing a considerable role, and bioacoustics monitoring was a significant field of study. Visual classification of animals is done by using either satellite images or established camera images. Nevertheless, due to some circumstances, image processing techniques cannot be applied. Then the acoustical classification techniques are taken place to encounter those problems. Even with acoustical methods, a remote observing method is required due to a few issues. Applying an IoT based acoustic classification system was designed using Convolutional Neural Networks (CNN), which is beneficial for those who are interested in monitoring ecosystems such as animal scientists, zoologists, and environmentalists. The hardware implementation was designed to collect the data from the place it was placed. The audio clips were preprocessed using the Mel-frequency Cepstral Coefficient (MFCC). A CNN architecture based on TensorFlow was used for the training process. To train and test the network, 400 sound clips of two seconds, such that 40 per each ten animal species, which were gathered from online libraries and formatted using Audacity, were used. The network was trained by changing the different gradient descent optimizers and eventually obtained the confusion matrices for each. The best result was gained by the AdaDelta, Gradient Descent, and RMSProp optimizers with 91.3% accuracy for each. Among them, AdaDelta had the most stable and increasing learning approach. As a future extension, to improve accuracy, a large number of data will be used.