Abstract
Tea diseases can significantly impact crop yield and quality, necessitating accurate and efficient recognition methods. This study presents WaveLiteNet, a lightweight model designed for tea disease recognition, addressing the challenge of inadequate disease feature extraction in existing approaches. By integrating 2D discrete wavelet transform (DWT) with MobileNetV3, the model enhances noise suppression and feature extraction through an adaptive thresholding strategy in the 2D DWT. The extracted frequency-domain features are fused with depth features from the Bneck structure, enabling a more comprehensive representation of disease characteristics. To further optimize feature extraction, a convolutional block attention module (CBAM) is incorporated within the Bneck structure, refining the network's ability to assign optimal weights to feature channels. A focal loss function also replaces traditional cross-entropy loss to mitigate sample category imbalance, improving recognition accuracy across varying distributions. Experimental results show that WaveLiteNet achieves a 98.70% recognition accuracy on five types of tea leaf diseases, with a parameter count of 3.16 × 10⁶. Compared to MobileNetV3, this represents a 2.15 percentage point improvement in accuracy while reducing the parameter count by 25.12%. These findings underscore WaveLiteNet's potential as a highly efficient and lightweight real-time crop disease recognition solution, particularly in resource-constrained agricultural environments.