# LightWaveNet: a lightweight wavelet-enhanced high-low-frequency-aware network with multi-stage supervision for rice disease recognition

## TL;DR

LightWaveNet is a lightweight and efficient deep learning model for rice disease recognition that balances accuracy and computational cost.

## Contribution

The novel contribution is a wavelet-enhanced network with multi-stage supervision for capturing high- and low-frequency features in rice disease recognition.

## Key findings

- LightWaveNet achieves 95.90% recognition accuracy with only 0.28 M parameters and 0.02 G FLOPs.
- It outperforms the lightest Mobilenetv2 model in both computational efficiency and accuracy.
- The model effectively captures fine-grained textures and structural patterns of rice diseases.

## Abstract

Accurate identification of rice diseases is critical for ensuring food security and advancing intelligent agricultural management. However, existing deep learning methods, while achieving high accuracy, often involve heavy computational costs and complex models, which limit their deployment on resource-constrained agricultural devices. More importantly, most of these methods rely on spatial domain representations and cannot model both high- and low-frequency information, making it difficult to capture fine-grained textures and overall structural features of diseased areas simultaneously.

To address these challenges, this study proposes a lightweight wavelet-enhanced high-low-frequency-aware network (LightWaveNet) for rice disease recognition. Specifically, LightWaveNet employs a parallel structure of wavelet convolution and max pooling to achieve collaborative learning of high- and low-frequency features, enabling effective extraction of both fine-grained textures and overall structural patterns. In the downsampling stage, a parallel design of max pooling and average pooling is adopted to further preserve the complementarity of frequency features. In addition, a multi-stage supervision mechanism is introduced to constrain and optimize features at different levels during training, thereby improving convergence speed and model robustness.

Experimental results demonstrate that LightWaveNet achieves a favorable balance between accuracy and efficiency. With only 0.28 M parameters and 0.02 G floating-point operations (FLOPs), it reaches 95.90% recognition accuracy. Compared with the lightest Mobilenetv2 model among the comparison methods (2.24 M parameters and 0.30 G FLOPs), LightWaveNet exhibits lower computational complexity while achieving higher recognition accuracy.

This study provides a feasible solution for rapid rice disease identification and intelligent prevention, while also offering new insights into the design of lightweight recognition networks for agricultural applications.