Application of real-time detection transformer based on convolutional block attention module and grouped convolution in maize seedling

Yunlong WuShouqi YuanYue TangLingdi Tang^*

• Jiangsu University, Zhenjiang, China

The intelligent detection and counting of maize seedlings constitute crucial components in future smart maize cultivation and breeding. However, the detection of maize seedlings in field environments faces substantial challenges due to their relatively small target size and the complex environment of the farmland. To achieve rapid and precise identification and counting of maize seedlings in the complex environment of large fields, this study proposes an improved detection model named CBAM-RTDETR. Based on the original feature extraction backbone network of RT-DETR, the model introduces the CBAM module and grouped convolution. While strengthening the shallow edge detail information of the seedlings and increasing the feature diversity, it effectively balances the accuracy and real-time performance of the model. Experimental results validated the research hypothesis, demonstrating that the CBAM-RTDETR model achieved a mean Average Precision at 0.5 IoU threshold (mAP0.5) of 92.9%, a mean Average Recall (AR) of 64.4%, and a Frames Per Second (FPS) of 87f/s on the test dataset, all of which are better than the comparison model.