Image Classification Experiment Based on Deep Learning Algorithm: Performance Optimization and Comparison

This paper optimizes the classification task of corn leaf disease images by combining the improved EfficientNet model with FPN (Feature Pyramid Network) and SE-Net (Squeeze-and-Excitation Networks), solves the shortcomings of traditional image classification methods in detail feature extraction and scale invariance, and improves classification accuracy. This paper uses the Plant Village public dataset, covering four categories: health, gray spot, rust, and leaf spot. This paper improves EfficientNet by applying FPN and SE-Net, enhancing its ability to handle multi-scale features and sensitivity to key disease features. FPN improves the network’s ability to capture targets of different sizes, while SE-Net optimizes the expression ability of features by adaptively adjusting channel weights. These two improvements make the model perform well in the task of corn leaf disease image classification. The improved EfficientNet outperforms ResNet-50, DenseNet, VGGNet (Visual Geometry Group Networks), GoogLeNet, and Xception deep learning models in various indicators, with an accuracy of 98.8%, a macro precision of 98.4%, a macro recall of 98.5%, and a macro F1 of 98.4%. These findings show that the improved EfficientNet can effectively improve the classification accuracy of corn leaf disease images, especially when processing multi-scale lesion features. In addition, ablation experiments show that the model’s performance is mainly due to the application of FPN and SE-Net.