A hybrid of residual network and inception neural network model for wildlife detection and identification

Abstract

Machine learning has significantly transformed various domains, with deep learning architectures playing a crucial role in computer vision applications. Convolutional neural networks (CNNs) have demonstrated remarkable success in image classification and object recognition tasks. However, traditional CNN architectures often exhibited limitations in handling complex feature extraction and generalization, particularly in wildlife identification where intra-class variations were high. The challenge in wildlife identification arose due to factors such as varying lighting conditions, occlusions, background clutter, and pose variations, which made it difficult for sinle model architectures to achieve high accuracy and robustness. This study sought to address these challenges by first designing and implementing individual Residual Network (ResNet) and Inception models to establish baseline performance, and then developing a hybrid ResNet-Inception model aimed at enhancing feature extraction, optimizing classification performance, and improving generalization capabilities in wildlife identification tasks. The Animals with Attributes 2 (AwA2) dataset was used to train and evaluate the models, and their performance was assessed using standard classification metrics, including accuracy, precision, recall, and F1-score. The WildlifeReID-10k dataset served as an external validation set. The hybrid approach leveraged ResNet’s ability to mitigate vanishing gradient problems through residual learning and Inception’s capability to capture multi-scale spatial features, thereby creating a more robust and efficient architecture. The results demonstrated that ResNet101 achieved an accuracy of 93.5%, Inception v3 achieved 95.6%, while the proposed hybrid model achieved 98%, confirming its superior performance in distinguishing visually similar species and enhancing generalization. The findings of this study provide a practical contribution to biodiversity conservation by enabling improved automated wildlife identification systems that support ecological monitoring, species recognition, and anti-poaching surveillance. By addressing the limitations of singlemodel approaches and demonstrating the advantages of hybrid deep learning architectures, the study sets a new benchmark in wildlife identification and reinforces the integration of artificial intelligence into environmental conservation practices.