SPQ: Similarity-Preserving Posttraining Quantization for Portable Crop Disease Detection

Abstract

The traditional method of diagnosing crop diseases through expert visual observations is expensive, time-consuming, and prone to errors. Deep neural networks (DNNs), specifically, convolutional neural networks (CNNs), offer promising alternatives, but are not easily accessible to local farmers, due to high resource requirements and reliance on network connectivity. Lightweight models that can operate on resource-limited devices and without network access are crucial to address this gap. Although quantization offers a promising solution for building lightweight models for crop disease detection, it often reduces quality. To address this issue, this article introduces a similarity-preserving quantization (SPQ) method to convert high-precision DNNs (particularly, CNNs in this article) into lower precision models while maintaining similar feature representations. SPQ ensures comparable activation patterns for similar crop images in both the original and quantized models. Experimental results using the MobileNetV2 and ResNet-50 backbones as base full-precision models demonstrate that SPQ improves memory footprint more than three times while preserving the detection performance, and show superior performance to the state-of-the-art quantization methods, especially when the target bit-widths are narrower.