Abstract:

[Objective] Precise detection of maize leaf diseases plays a pivotal role in safeguarding maize yields and promoting sustainable agricultural development. However, existing detection algorithms often fall short in effectively capturing the intricate morphological details and shape characteristics of disease spots, particularly under challenging scenarios involving small disease targets. To overcome these challenges, a novel maize leaf disease detection algorithm, DCC-YOLOv10n, is presented in this research, which is specifically optimized for scenarios involving small-scale disease targets. [Methods] The maize of the proposed method lay in three innovative architectural enhancements to the YOLOv10n detection framework. Firstly, a DRPAKConv module was designed, which built upon the Arbitrary Kernel Convolution (AKConv). DRPAKConv replaced the conventional 3×3 convolutions that typically occupied a large proportion of the model's parameters. It featured two parallel branches: a dynamic sampling branch that adjusted the sampling shapes based on the spatial distribution of disease patterns, and a static convolution branch that adapted kernel sizes to retain spatial coverage and consistency. This design significantly enhanced the network's capability to recognize small-scale disease spots by dynamically modulating the receptive field and focusing on localized lesion details. Secondly, an improved feature fusion part was introduced by replacing the traditional C2f feature fusion module with a novel CBVoVGSCSP module. This redesigned module aimed to address the issue of gradient vanishing in deep feature fusion networks while reducing computational redundancy. CBVoVGSCSP preserved rich semantic information and improved the continuity of gradient flow across layers, which was critical for training deeper models. Furthermore, it enhanced multi-scale feature fusion and improved detection sensitivity for lesions of varying sizes and appearances. Thirdly, the Convolutional Attention-based Feature Map (CAFM) was incorporated into the neck network. This component enabled the model to effectively capture contextual relationships across multiple scales and enhanced the interaction between spatial and channel attention mechanisms. By selectively emphasizing or suppressing features based on their relevance to disease identification, the module allowed the model to more accurately distinguish between diseased and healthy regions. As a result, the model's representational capacity was improved, leading to enhanced detection accuracy in complex field environments. [Results and Discussions] Extensive experiment was conducted on a specialized maize leaf disease data set, which included annotated samples across multiple disease categories with diverse visual characteristics. The results demonstrated that the DCC-YOLOv10n algorithm outperformed baseline models and several state-of-the-art detection frameworks. Compared with YOLOv10n, the optimized algorithm demonstrated a reduction in computational complexity by 0.5 GFLOPs, with the model parameters compressed to merely 2.99 M. Significant improvements were observed in precision, recall, and mean average precision, which increased by 1.7, 2.6, and 1.7 percentage points respectively, and achieved 96.2%, 90.3%, and 94.1%. The findings underscored the robustness and adaptability of the DCC-YOLOv10n algorithm under challenging conditions. [Conclusions] The DCC-YOLOv10n algorithm presents a significant advancement in the field of agricultural disease diagnostics by addressing the limitations of existing methods with respect to small-target detection. The novel architectural components—DRPAKConv, CBVoVGSCSP, and CAFM integrated with attention fusion—not only significantly enhance the model's detection performance, but also advance the development of intelligent, data-efficient, and highly accurate disease monitoring systems tailored for modern agricultural applications. This research would serve as a valuable reference for future developments in lightweight, efficient, and accurate maize disease detection models, and offered practical significance for intelligent maize management.

Key words: maize leaf, disease detection, small target, DCC-YOLOv10n, AKConv, CAFM