CTRNet: a lightweight and efficient deep learning model for field maize whorl identification.

Due to the extremely small target size of maize whorls at the whorl stage, frequent leaf occlusion, and the presence of complex backgrounds and dynamically varying illumination in field environments, accurate detection remains highly challenging. To address these issues, this paper proposes a maize whorl detection network termed the Contextual and Texture-enhanced Representation Network (CTRNet), which enhances feature representation through multi-scale contextual interaction and fine-grained feature modeling. Specifically, a multi-scale contextual interaction module (MSCIM) is introduced across four feature scales to improve semantic modeling under occlusion conditions. A dual-channel fine-grained feature enhancement module (DCFEM) is further integrated to strengthen detailed representations of small-scale targets, while a gated adaptive information fusion module (GAIFusion) is employed to efficiently aggregate multi-scale features. In addition, an adaptive channel-edge attention mechanism (ACEA) is incorporated into high-level features to suppress background interference and mitigate the effects of illumination variations. Experimental results demonstrate that, under complex field scenarios, CTRNet improves the mAP@0.5 from 81.6% to 84.7% with only 2.38 M parameters, significantly enhancing detection robustness for small targets under occlusion and varying lighting conditions. This study provides an efficient and lightweight technical solution for precision monitoring and pesticide application in maize pest management.