An efficient deep learning based disease detection method for detection of multiple lesions in tea leaves

An automated disease detection method is crucial for early detection of diseases in tea plants to ensure adequate and quality tea production. Existing studies on automated disease detection methods left a wide range of scope to improve detection accuracy and computational latency. Moreover, studies on the localization of multiple disease lesions in tea leaves have not been attempted to date. This study proposes an efficient and automated deep learning-based detection method, based on the improved architecture of the EfficientDetD0 model to accurately locate the lesions in tea leaves. The model incorporates a new feature enhancement module that replaces the squeeze and excitation network in the backbone to reduce parameter redundancy in the network. In addition, a contextual module is integrated between the backbone and the Bidirectional Feature Pyramid Network (BiFPN) of the model to enhance the feature representation ability of the model. Furthermore, the number of BiFPN layers and the number of convolutions in the Bounding Box/Class prediction network are tuned in the model architecture to achieve optimized results. The proposed model outperformed other state-of-the-art detectors, achieving a mean average precision value of 84.5% with a comparable GPU latency.