<p>Crop fertilization, pesticide spraying and weed treatment are crucial links in the process of crop growth, and the effective implementation of these tasks depends on accurate crop and weed identification. The recognition technology based on vision can realize automatic recognition. In this study, a new corn–weed detection model named YOLOv8–GAS-based YOLOv8 was proposed to address the widespread problems of poor recognition performance, difficulty in effectively dealing with complex scenarios, and difficulty in co-existing lightweight detection models and recognition accuracy. First, the newly developed GRCSPESIN (ghost reparameterization cross-stage partial connections stage integration network) module was introduced and integrated into the C2f (channel-to-pixel) feature extraction layer in YOLOv8. Then in the neck of the network, the multi-scale feature extraction by sharing convolution kernels of SPC (Shared Pyramid Convolution) module was introduced and the context-guided feature fusion module, ACFM (Adaptive Context Fusion Module) was employed. Finally, a new corn–weed data set was constructed based on an image acquisition of a complex unmanned farm maize test field, which was used for a comparative analysis of YOLOv8–GAS and the baseline model. The results of comprehensive evaluation of the obtained data set demonstrate that the proposed model has excellent performance, with a 3.2% increase in mAP@0.5, a 10.3% decrease in model parameters, and a 16.05% decrease in calculation amount compared with YOLOv8n.</p>