Abstract: To address the low accuracy of pitaya ripeness recognition caused by complex backgrounds and target occlusions in unstructured environments, this study proposed an improved YOLOv8n-based model for pitaya ripeness detection. The YOLOv8-MD model was built based on YOLOv8n by integrating the MobileNetV3-Small backbone network and incorporating the DIoU loss function. This design fused local low-dimensional information and global features of pitayas in field images, thereby enhancing the model’s capability to extract ripeness-related features and mitigating challenges associated with target occlusion and small-sized targets in unstructured settings. Experiment results demonstrated that in unstructured orchard environments, the proposed model achieves a precision of 94.2%, a recall of 87.5% and a mean average precision (mAP) of 95.4%, representing improvements of 7.4%, 7.0%, and 6.1% respectively over the baseline YOLOv8n model. Concurrently, the model’s parameter count was reduced to 2.77 M, and the total floating-point operations (FLOPs) were compressed to 6.4 GFLOPS, a 20.9% reduction compared to the baseline. Deployment testing on the Nvidia Orin Nano platform yielded an inference speed of 83.13 FPS, which was 9.2% to 41.1% faster than the traditional YOLO models. This research provided a valuable framework for the development of field-based ripeness detection models for pitayas and other fruits with similar characteristics.