Cat and Dog Breed Classification Based on SE-DenseNet Integrated Modeling
Shitao Zhang, Yuanming Wang, Fangfang Sun, Yong Yang · 2024
With the rapid expansion of the pet market, the increasing diversity and complexity of pet species have introduced novel challenges in the domain of pet identification. This study proposes an enhanced model for the classification of cat and dog breeds, utilizing the DenseNet architecture. The model integrates DenseNet's feature reuse mechanism with the attention-enhanced Squeeze-and-Excitation (SE) block to augment the efficiency and accuracy of feature extraction. To broaden the scope and diversity of model training, the study employed several publicly available datasets, supplemented by additional cat and dog image data collected through the network. During the preprocessing phase, this study employs YOLOv10 for Region of Interest (ROI) identification to eliminate background and extraneous elements, thereby mitigating the adverse effects of interfering information on model training and inference. The experimental results demonstrate that, with a model size of merely 36.5 MB, the model attains classification accuracies of 89.20% on the validation set and 82.96% on the test set across 37 categories in cat and dog breed classification tasks. These findings underscore the model's potential for applications in cat and dog breed classification.