Model compression using post-training dynamic range quantization for plant disease recognition

This study investigates model compression techniques for plant disease recognition using leaf images, focusing on reducing the computational burden of deployment on resource-constrained devices. Utilizing the PlantVillage dataset, the study trained and evaluated 11 prominent CNN architectures by employing a pocket algorithm for performance optimization. This study examines three training scenarios: random weight initialization, ImageNet pre-trained weights with trainable layers, and ImageNet pre-trained weights with non-trainable layers. The results demonstrated superior performance in scenarios with trainable layers, with the Xception model achieving the highest minimum accuracy. Subsequently, the study implements magnitude-based pruning using the PolynomialDecay strategy, successfully reducing the model size by 90% while maintaining comparable accuracy. Further compression was achieved through post-training dynamic range quantization, yielding a compression ratio exceeding 90% for most models without significant accuracy degradation. This study highlights the effectiveness of combining pruning and quantization to deploy accurate and efficient plant disease recognition models on devices with limited resources.