A dual-branch network integrating spatial domain and frequency domain information for detecting tea leaf blight at different stages

Abstract

Tea leaf blight (TLB) is a common tea disease, and accurate detection of the different stages of TLB helps in tea disease control. The color and shape of TLB spots at different stages vary greatly and are easily confused with complex backgrounds; hence, the accuracy of existing methods for detecting TLB at different stages is not high. In this study, a dual-branch TLB detection network (DBTDNet) combining spatial domain and frequency domain information was designed for the accurate detection of TLB at different stages. The dense depthwise separable (DDS) module and wavelet-based feature extraction (WBFE) Bottleneck were introduced into the spatial feature extraction (SFE) branch and frequency feature extraction (FFE) branch of DBTDNet to extract spatial domain features and frequency domain features, respectively, and enhance the localization and recognition of TLB spots at different stages. A multiscale wavelet transform convolution (MSWTC) module was also added to the FFE branch to separate the multi-scale frequency information and obtain clearer shape and texture features of TLB spots. A linear layer was introduced between the dual-branch structures to reduce the gradient information loss. In addition, to better capture TLB spots of different sizes, this study designed a multidimensional neural network (MNNet) structure in the feature fusion part of DBTDNet for fusing the information of different scale feature maps from the output of the dual branch. The experimental results showed that the proposed DBTDNet could more accurately detect TLB spots at different stages than the existing state-of-the-art network models. The mAP@0.5 values of its detection results for yellow TLB spots in the early stage, white TLB spots in the middle and late stages, and total TLB spots were 74.5%, 75.3%, and 75%, respectively, which were 13.2%, 7.5%, and 10.5% higher, respectively, than the baseline model YOLOv9 detection results.

Plant phenotyping relevance

茶葉の病斑という植物の病害状態を画像から検出・認識する深層学習手法を開発し、既存モデルと性能比較しているため、植物フェノタイピング手法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.