Recognizing tea diseases with fusion on 2D DWT and MobileNetV3

Diseases have posed the serious threaten on the yield and quality of tea production. An accurate and rapid recognition of leaf diseases is essential to the instant diseases prevention of tea plantation. Deep learning can be expected to realize a rapid and accurate identification of tea diseases in natural environment with the advantages of low cost and high efficiency, compared with typical disease diagnosis. However, the previous models have much more parameters and computational complexity for the leaf diseases diagnosis. Furthermore, the lightweight models cannot fully meet the fine-grained feature extraction. In this study, a disease recognition network (CBAM-TealeafNet) was proposed to extract the frequency features by the 2D discrete wavelet transform (2D DWT) and depth features by the bneck structure. Frequency features were then decomposed to suppress the high-frequency components. The fused feature module was used to reduce the impact of noise on the features for the features enhancement. CBAM (convolutional block attention module) was embedded to improve the feature extraction capability in the bneck structure. The weights were allocated into the feature channels and spatial position features of diseases. The function of focal loss was employed to replace the primitive cross-entropy loss, in order to better resolve the imbalance influences on sample class for the high accuracies. Totally, 3, 260 disease images of Shaanxi Tea No.1 and Longjing No.43 were captured, including five tea disease categories: gloeosporium theae-sinensis miyake, colletotrichum camelliae massee, cercospora theae breadade haan, exobasidium vexans masse, and phyllosticta theicola petch. The real environment was also simulated to evaluate the datasets. The images were then enhanced. Experiments were carried out to validate the optimal model structure and the improvement analysis of each component. The model was optimized for the hyperparameters setting. The final optimal learning rate was 0.000 5, which was derived from an initial learning rate range of 0.000 05-0.005. In addition, the whole recognition structure and the base model structure of MobileNetV3 were optimized to determine the optimal number of fusion layers and the fusion ratio on frequency and depth feature channels. The results showed that the CBAM-TealeafNet model was achieved in the higher accuracy on the tea disease recognition, compared with the previous models. The number of parameters was ranked secondly last only to MCA-MobileNet model. The CBAM-TealeafNet model increased the accuracy by 2.15%, whereas, the number of parameters decreased by 25.12%, compared with the NobiNetV3. Misidentification images and confusion matrix indicated that the CBAM-TealeafNet shared the better performance to highly distinguish between foreground and background, thus greatly improving the situation of disease confusion. In addition, the functions of cross-entropy and focal loss were compared to verify the accuracy of recognition on the dataset imbalance. Moreover, the CBAM model performed the superior to the SENet and ECANet, in terms of performance improvement. The CBAM-TealeafNet was employed to recognize the tea diseases. An accuracy of 98.70% and a F1-Score of 98.69% were achieved with the parameter number of 3.16×10⁶ and FLOPs (floating-point operations) of 4.5×10⁸. The CBAM-TealeafNet can be expected to effectively identify the diseases under the complicated environment, particulary with the characters of less parameter memory and higher inference speed. Misidentification of CBAM-TealeafNet will be reduced in future investigation. This finding can also provide a strong reference for the model construction on the recognition of common tea leaf diseases.