Detection of Tea Leaf Blight in Low-Resolution UAV Remote Sensing Images

The accurate detection of tea leaf blight (TLB) from low-resolution (LR) images of tea plants is a challenging task. The small size, high density, and blurred edge details of diseased tea leaves in the LR images from remote sensing with an unmanned aerial vehicle (UAV), as well as the complex backgrounds in the images and the similarity between the color and texture features of TLB spots and the background, cause state-of-the-art methods to give low detection accuracy. This study proposes a two-stage network that incorporates a super-resolution (SR) network RFBDB-GAN with a lightweight detection network (LWDNet) for the accurate detection of TLB in LR UAV images. RFBDB-GAN is applied to recover the details of UAV images. A multiscale two-stage degradation is introduced to simulate the degradation of UAV images caused by the use of various shooting heights to reduce the solution space of RFBDB-GAN. RFB dense residual block (RFBDB) and spectral normalization (SN) term are added to the generator and discriminator, respectively, to recover the texture details and make the TLB spots more obvious. An SIoU loss and an enhanced lightweight PANNet (ELPNet) are employed in LWDNet to achieve accurate TLB detection using a model with a very small size. Experimental results show that RFBDB-GAN yields noticeable improvements over other SR methods, and the model size of LWDNet is 716 kB or 5% of the size of the baseline YOLOv5s. Compared with other state-of-the-art methods, the proposed method has a higher detection accuracy for TLB in LR UAV remote sensing images.