Pepper-Det: A real-time framework for automated chili pepper detection and harvesting point localization

The pepper harvesting robot is crucial for reducing labor intensity and improving crop production efficiency. In complex environments, fruit recognition and picking point localization are major challenges. This paper proposes Pepper-Det for real-time pepper detection and picking point localization. Built on YOLOv11, it replaces traditional convolution with dynamic perceptual adjustment and channel-space attention, enhancing feature modeling in varying scales and occlusion. We combine an efficient channel attention module with Mobile Inverted Bottleneck Convolution (MBConv) to form the EffectiveSE-MBConv (EMBC) module, reducing computation without sacrificing performance. Additionally, the decoupled detection head (LADH) separates object and picking point tasks, improving feature utilization and reducing model parameters. Experiments on a self-built pepper dataset show Pepper-Det achieves an object detection mAP@50 of 85.1%, a keypoint detection mAP@50 of 88.7%, and an overall mAP of 83.0%. Its computational complexity is 5.4 GFLOPs, with a 2D pixel Euclidean distance error of 11.13 pixels and a median 3D Euclidean distance error of 9.51 mm. These results outperform other models in the experimental comparison. Cross-dataset validation on the artificial chili pepper model (ACPM) dataset and eggplant datasets demonstrates that the model still achieves state-of-the-art performance, with mAP@50 reaching 95.6% and 94.3%, respectively, and achieved a 95% success rate in the simulated pepper-picking task in the laboratory. This research effectively addresses the technical challenges faced by the visual perception environment of intelligent fruit-picking robots and provides technical support for the visual perception technology of harvesting robots.