A hybrid deep learning based robust framework for cattle identification

Abstract

This study proposes a deep learning-based framework for recognizing cows based on images of their muzzle, and faces. This method works well when dealing with missing or false insurance claims. This study proposes a hybrid multi-stage framework consisting of different phases such as augmentation, denoising, enhancement, and classification. The proposed framework is developed by hybridizing convolutional denoising autoencoders (CDAE), least squares generative adversarial network (LS-GAN), Xception feature extractor, and a convolutional neural network (CNN). CDAE is used to initiate the process of denoising noisy images. LS-GAN is used to improve the characteristics of denoised images by enhancing the image by elimination of the residual noise. The Xception is utilised to extract significant and optimal features, and CNN is then used for classification. Various comparative methodologies are used to assess the proposed approach at different phases through several statistical measures. The proposed framework achieved 97.27% accuracy using the test datasets, which is higher than the comparative approaches.