An accuracy—efficiency trade-off study of lightweight CNNs for plant leaf disease classification

Abstract

Plant leaf disease classification is important for early diagnosis and crop management, but deploying Convolutional Neural Network (CNN) models in resource-constrained settings remains limited by memory and latency constraints. This paper presents Latency-CNN, a lightweight architecture that combines early spatial downsampling, depthwise separable convolutions, and a compact Global Average Pooling (GAP)-based classifier. Rather than introducing a new convolutional operator, this study evaluates a deployment-oriented configuration of existing lightweight components under practical CPU efficiency constraints. On a 15-class PlantVillage-derived dataset, Latency-CNN achieves 95.98% accuracy with 107K parameters, a 0.41 MiB FP32 size, 0.156 GFLOPs, 0.078 GMACs, and 1.90 ms/image CPU latency. TensorFlow Lite dynamic-range quantization further reduces the model size to 0.136 MiB and latency to 1.56 ms/image while maintaining comparable accuracy at 96.07%.Additional component-level ablations show that replacing depthwise separable convolutions with standard Conv2D improves accuracy to 97.81% but increases computation and latency, while replacing the compact classifier with a Flatten-based classifier substantially increases model size. Synthetic perturbation results show sensitivity to low illumination, with accuracy dropping to 76.04% under 0.65 × brightness. These results suggest that Latency-CNN is a possible lightweight configuration for accuracy-efficiency trade-off analysis under controlled evaluation conditions, while illumination-aware training and real-field validation remain necessary before deployment claims can be made.