Leaf Classification for Local Area Network Deployment: Comparison and Pruning of Resnet18 and Attention Modules (SE/CBAM)

Abstract

This paper focuses on resource-constrained edge deployment for the 15-class grayscale image classification task using the Swedish Leaf dataset. It systematically compares the standard ResNet18, the channel attention SE (Squeeze-and-Excitation) model, and the channel and spatial attention CBAM model. Structured pruning and distillation fine-tuning are then applied to the best-performing model. The data is divided into train, verification, and test sets, with class balance, each comprising 55, 10, and 10 images per class (totalling 825/150/150 images, respectively). Under a unified training strategy (frozen backbone, AdamW, label smoothing, no Mixup), SE achieved a test accuracy of 0.9133, outperforming the Baseline (0.8733) and CBAM (0.8667). Paper further performed progressive channel pruning on stages 3 and 4 of the SE models (6% per step, over three steps, resulting in total reduction parameters of approximately 23.3% and 19.1% FLOPs). This was followed by low-learning-rate, long-cycle fine-tuning. The final test accuracy was 0.8933, which remains higher than the Baseline despite significant model compression. Using Grad-CAM, the paper quantified interpretability through three metrics: leaf coverage, leaf IoU, and leaf margin IoU, with the following results: CBAM > SE > Baseline, indicating that attention modules focus more on the leaf regions. Experiments also showed that eight-view TTA significantly reduced accuracy in this setup, suggesting TTA can have side effects when mismatched with data augmentation/ normalisation. This paper presents comprehensive training/validation curves, confusion matrices, interpretability metrics, and the pruning-accuracy trade-off, providing a reproducible reference for the engineering implementation of lightweight plant phenotyping classification.