Navigation line detection can serve as a pivotal procedure in the visual navigation of agricultural machinery. Driving direction or path guiding of agricultural machinery can be accurately extracted from field images in real time. Instance segmentation has been mostly used for this task at present. Crop row edges can be extracted to fit navigation lines after complex post-processing. However, cumbersome workflows and large computational overhead have severely restricted their application under real-time agricultural navigation. It is often needed to have high efficiency and instant response. This study aims to propose an end-to-end navigation line regression model, called Keypoint Plant Line Network (KPLNet). Navigation line detection of agricultural machinery was realized with superior precision and efficiency. Different from conventional instance segmentation, the KPLNet framework directly represented navigation lines in the form of keypoints for prediction. Post-processing was removed to improve inference speed and detection accuracy. Firstly, the Central Attention Mechanism (CAM) was constructed in the architecture of KPLNet using a geometric prior. Contextual integration of features was concentrated at key positions. Accurate keypoint positioning of navigation lines was realized to effectively enhance the extraction and efficiency of critical features. Secondly, a streamlined feature fusion structure, named Single Feature Fusion (SFF) was proposed, according to the small target scale variation and the demand for single-target detection only in the navigation line detection task. SFF structure effectively extracted navigation line features. While the network structure was improved, the real-time performance of the model was verified. Redundant calculation and parameter costs were reduced for the edge deployment requirements of agricultural machinery. Finally, an improved Dual-Optimization Particle Swarm Optimization (D-PSO) was introduced for hyperparameter optimization during training. The high accuracy of the model was obtained. A better balance between performance and computational efficiency was achieved after optimization. Experiments were conducted on the maize and peanut-soybean intercropping field. The KPLNet was also compared with mainstream instance segmentation, including YOLOv8n-seg and YOLO11n-seg. The experimental results show that KPLNet performed the best in terms of detection accuracy and real-time performance. Concretely, the navigation line endpoint detection accuracy of KPLNet reached 99.5% in maize and peanut-soybean intercropping fields, and the heading angle deviations were as low as 0.79° and 0.73°, respectively, indicating extremely high detection precision. In terms of inference efficiency, the lower average inference time of KPLNet was obtained to eliminate the post-processing time consumption, with the maximum frame rate of 119.2 frames per second, fully meeting the real-time navigation requirements of agricultural machinery. Ablation tests verified that the CAM, SFF, and D-PSO module contributed to the accuracy and inference speed. The experiments validated that the end-to-end navigation was superior to the simple workflows and had low computational overhead in the task of navigation line detection, compared with the conventional instance segmentation. This finding can provide reliable technical support to develop efficient and stable navigation for agricultural machinery. A technical reference can also offer guidance for the intelligent and precise operation of agricultural equipment in complex planting environments in smart agriculture.

An efficient and accurate detection is often required for the grape maturity during harvesting in precision agriculture. This study aims to propose a lightweight, high-precision, and rapid detection for the grape fruit maturity using an improved YOLOv7 model, referred to as YOLOv7MCA. The YOLOv7 model was modified to introduce the MobileNetV4 as the backbone network of feature extraction. The MobileNetV4 was utilized to reduce the number of parameters. In turn, the computational load was significantly reduced to maintain the high detection accuracy and efficiency. Computational cost and performance of the model were balanced suitable for real-time applications with the limited computational resources. In addition, Convolutional Block Attention Module (CBAM) was incorporated to enhance the feature extraction of the YOLOv7 neck part. CBAM was introduced the attention mechanisms in both the channel and spatial dimensions, in order to focus on the more relevant features while suppressing less important ones. The attention mechanism was significantly improved to select the discriminative features, thereby enhancing the accuracy of grape maturity detection, especially in the complex or cluttered images. Furthermore, Adaptive Spatial Feature Fusion (ASFF) module was introduced into the feature fusion. The head part of the YOLOv7 model was optimized after the adaptive fusion of multi-scale feature maps. As such, the self-adaptive fusion mechanism was effectively detected both small and large objects, suitable for the various targets of different sizes. The precision and efficiency were improved among different types of images. A wider range of object scales was determined in real-world agricultural applications. Experimental results demonstrate that the improved YOLOv7MCA model was achieved in a precision of 95.2%, a recall rate of 87.2%, and a mean average precision of 93.9% on the grape image test set. The average detection time was 52.2 ms, and the memory usage of the improved model was 53.6 MB. The performance of object detection was outperformed the existing models, including Faster R-CNN, SSD, YOLOv5, YOLOv7, YOLOv8n, YOLOv9t, and YOLOv10n. The improved YOLOv7MCA model was reduced the memory usage to maintain the high detection accuracy with the less detection time. The speed and memory efficiency were especially important for the practical deployment on embedded devices with the limited resources. The improved model was suitable for real-time applications, such as automated grape harvesting. The rapid and accurate object detection was realized for the grape harvesting in agricultural applications. The potential of the YOLOv7MCA model was highlighted to enhance the effectiveness and scalability of automation systems in precision agriculture.