Research on the Application of EfficientNet-B4 Model Enhanced with Attention Mechanism in Animal Image Classification

Animal image classification is widely used in ecological monitoring, pet identification, and wildlife protection. However, traditional convolutional neural network models (such as ResNet and standard EfficientNet) often lack targeted capture of key features (e.g., animal fur texture and body contour), leading to limited classification accuracy under complex backgrounds or pose variations. To address this issue, this study proposes an SEEnhancedEfficientNetB4 model, which embeds a Squeeze-and-Excitation attention module into the EfficientNet-B4 backbone. The SE module adaptively adjusts the weight of each feature channel to enhance the extraction of discriminative features for animal images. A two-stage transfer learning strategy is adopted: first, freeze the backbone network to train the classification head, and then unfreeze part of the backbone for fine-tuning.