Abstract:

[Objective] Kiwifruit leaf diseases are difficult to identify automatically because different diseases can exhibit highly similar visual characteristics, while lesion size, shape, and distribution vary considerably under natural growth conditions. In particular, small and early-stage lesions are frequently missed by conventional object detection models. Manual disease diagnosis is also labor-intensive and strongly dependent on the experience of agricultural technicians. Therefore, this study aims to develop an accurate and lightweight deep-learning-based method for automatic kiwifruit leaf disease detection, improving the recognition of small lesions while reducing model parameters and computational complexity. [Methods] A kiwifruit leaf disease dataset was constructed using images collected from a kiwifruit production orchard in Deyang, Sichuan province, China. Three common diseases, namely bacterial canker, brown spot, and anthracnose, were selected as detection targets. After annotation, 25 453 lesion instances were obtained, with an average of 10.47 lesions per image. The dataset was divided into 2 136 training images, 147 validation images, and 147 test images. To address the limited size of the training set, offline data augmentation was applied exclusively to the training images. The training set was expanded to 3 372 images, whereas the validation and test sets remained unchanged. Based on RT-DETR-r18, an improved model named RT-DETR-DPF was developed through three key modifications. First, the BasicBlock in the backbone was replaced with a detail-enhanced convolution (DEConv) module, which enhances the extraction of lesion edges, textures, and fine-grained structural information through multi-branch differential convolution. Second, the attention-based intra-scale feature interaction (AIFI) module was replaced with a polarity-aware linear attention (PolaFormer) module to strengthen the discrimination of disease-related regions through polarized query-key interactions. Third, a feature-focused diffusion pyramid network (FDPN) was designed to replace the conventional feature pyramid network-path aggregation network (FPN-PAN), improving the interaction and fusion of multi-scale lesion features. The model was further lightweighted by reducing feature dimensions, repeated blocks, and decoder queries. [Results and Discussions] Ablation experiments showed that DEConv, PolaFormer, and FDPN produced complementary improvements in detection performance. The complete RT-DETR-DPF model achieved a precision of 93.1%, a recall of 85.7%, an mAP50 of 92.3%, and an mAP50:95 of 75.6%. Compared with the baseline RT-DETR-r18, precision and mAP50 increased by 5.0 and 2.6 percentage points, respectively. Meanwhile, the parameter count decreased from 19.9 M to 13.2 M, representing a reduction of approximately 33.7%, while GFLOPs decreased from 56.9 to 37.7, corresponding to a reduction of approximately 33.7%. The inference speed was 67.0 f/s. Comparative experiments with other detection models showed that RT-DETR-DPF achieved higher precision than the evaluated RT-DETR variants, RetinaNet, YOLOv8m, YOLOv11l, YOLOv12m, and faster region-based convolutional neural network (Faster R-CNN), while maintaining relatively low model complexity. Compared with YOLOv11l, RT-DETR-DPF improved precision by 4.4 percentage points while reducing the parameter count and computational complexity by 47.8% and 56.5%, respectively. Compared with Faster R-CNN, RT-DETR-DPF improved mAP50 by 11.5 percentage points and reduced the parameter count by 68.0%. Feature visualization further showed that DEConv generated stronger responses around lesion boundaries and texture regions, PolaFormer produced more concentrated responses in disease-related areas, and FDPN enhanced feature interaction across different scales. Although the inference speed was slightly lower than that of the baseline model, the substantial reductions in parameter count and GFLOPs demonstrate a more favorable balance between detection accuracy and computational cost. [Conclusions] The proposed RT-DETR-DPF achieves a coordinated balance between detection accuracy and model complexity through complementary improvements in feature extraction, attention modeling, and multi-scale feature fusion. These results indicate the potential for accurate kiwifruit leaf disease detection in computationally constrained environments. Nevertheless, the current dataset exhibits limited diversity in geographic regions and environmental conditions, and the model requires further validation across different kiwifruit varieties and extreme weather conditions.

Key words: kiwifruit leaf disease, object detection, RT-DETR, attention mechanism, deep learning, lightweight model