A comparative analysis of the efficiency between different datasets in the identification of dogs and cats in a CNN

Abstract

Nowadays, several methodologies and algorithms are being developed to improve image recognition and increase efficiency in species identification. In this context, this paper introduces a comparative analysis of the efficiency between different datasets in the identification of dogs and cats in a convolutional neural network (CNN). In this study, a convolutional neural network architecture with 13 layers was evaluated using three different datasets. In the first set, the 'catsvsdogs' database from TensorFlow was used. In the second CNNN, the network was trained using a set of images that included only dog2 and cat2 species. Finally, in the third CNN the network was trained using a set of images that included only dog3 and cat3 species. The hypothesis put forward is that training a convolutional neural network with customized images of specific dogs and cats improves the accuracy in identifying these species compared to using the TensorFlow dataset. The performance of both models was evaluated using standard machine learning metrics. The results show that the accuracy of the convolutional neural network trained with personalized images increased significantly compared to previous results. Specifically, the recognition accuracy of specific dogs and cats improved considerably. In addition, the training time was reduced by approximately 94.8%, from 116 minutes to only 6 minutes. In conclusion, the use of personalized images in the training set can significantly improve the accuracy in identifying these species in a convolutional network, which can be especially useful in applications such as automatic pet feeders, where high accuracy is required when identifying the pet and providing the correct food.