Cnns Optimization for Corn Leaf Disease Detection Using Post-Training Quantization Techniques

Abstract

This paper proposes the use of post-Training quantization techniques, Float16 and dynamic range quantization, to optimize convolutional neural networks (CNNs) for corn leaf disease detection using a corn leaves subset of the PlantVillage dataset. Four CNN models are evaluated: VGG16, DenseNet121, MobileNetV2, and NASNet Mobile. The quantization methods reduce model by size up to 75%, while maintaining accuracy above 97%. DenseNet121 achieved the best performance of \mathbf{9 8. 7 0 \%} accuracy after dynamic range quantization. MobileNetV2 showed the most significant improvement in inference latency, decreasing from 146.28 ms to 27.14 ms per image after Float16 quantization. These results show the effectiveness of post-Training quantization techniques in optimizing CNN models with the potential for implementation in real-Time agriculture use.