ICTNet: Independent CNN–Transformer Dual-Branch Network for Poppy Object Detection

UAV patrol plays a crucial role in combating illegal poppy cultivation. However, the complex and diverse environments in which poppies are cultivated demand effective local and global feature modeling to improve detection performance while keeping computational costs low. Existing CNN–Transformer methods often rely on complex feature interaction to exploit their complementary strengths, and the scarcity of publicly available poppy datasets further restricts the development of poppy detection methods. To address the above issues, we construct an Independent CNN–Transformer Dual-Branch network (ICTNet), which separately models the CNN branch and the Transformer branch, thereby avoiding complex feature interaction. For the Transformer branch, we design a Gated Transformer Block (GTB). GTB employs gated linear attention to model complex spatial information with linear computational complexity, achieving a favorable balance between detection performance and computational efficiency. For the CNN branch, we design a Progressive Convolution Block (PCB) with an expanding receptive field, which effectively models local features while capturing rich contextual information. Moreover, we create a multi-scene poppy detection dataset named MS-Poppy, which contains 2443 images covering six UAV-view scenes. Extensive experiments are conducted on the MS-Poppy and the public GWHD 2021 dataset. Experimental results indicate that, compared with the advanced YOLO12, ICTNet reduces parameters by 11% while improving mAP@0.5 by 5.4% on the MS-Poppy dataset and by 1.0% on the GWHD 2021 dataset, respectively.