Abstract: Yellow dragon fruit (Selenicereus megalanthus) flowers open mainly at night, and their detection for automated pollination is affected by weak illumination, dense flower distribution, branch occlusion, and visually similar flower stages. To support real-time visual perception for pollination robots under field conditions, this study proposed a lightweight low-light yellow dragon fruit flower detection model, named Lightweight Low-light Dragon Fruit Flower YOLO11n (LLDF-YOLO11n), based on the YOLO11n framework. A multi-scenario image dataset was first constructed from a commercial yellow dragon fruit planting base in Saima Dam, Lujiang Town, Baoshan City, Yunnan Province, China. A total of 1988 images were collected under three representative illumination conditions, including supplemental array lighting, extremely low-light, and natural light. The flower targets were divided into three categories, namely bud, blossom, and wilted flower, covering the main flowering stages relevant to pollination management. All images were center-cropped at a one-to-one aspect ratio and resized to 720 × 720 pixels to standardize image scale and reduce irrelevant background interference. To improve annotation reliability under different illumination conditions, an illumination-specific labelling strategy was adopted. Images collected under supplemental array lighting and natural light were annotated directly using minimum bounding boxes. For extremely low-light images, brightness-assisted labelling was used. The original images were first processed with brightness and exposure compensation to make target contours more visible; after manual annotation, the label coordinates were mapped back to the original low-light images. After removing blurred and highly duplicated samples, 1600 images were retained as the basic dataset. These images included 524 supplemental-light images, 530 extremely low-light images, and 546 natural-light images. The dataset was divided into training, validation, and test sets at a ratio of 7:2:1. Horizontal flipping and random rotation were then applied to augment the dataset to 4800 images, including 3360 training images, 960 validation images, and 480 test images. The proposed LLDF-YOLO11n model was designed to address three main constraints in yellow dragon fruit flower detection: low-light image degradation, occlusion caused by dense flower clustering, and limited computational resources on edge devices. First, a Self-Calibrated Illumination (SCI) module was inserted into the first layer of the backbone network to calibrate low-light inputs before feature extraction. The SCI module was pretrained using unlabeled low-light flower images and then frozen during detector training, allowing the detection network to benefit from illumination-corrected inputs while preserving the enhancement capability of the pretrained module. Second, the original attention structure was replaced with a C2PSA_LSKA module, in which Large Separable Kernel Attention (LSKA) was introduced to model long-range spatial dependencies. This design decomposes large convolution kernels into separable one-dimensional depthwise convolutions and dilated depthwise convolutions, thereby enlarging the effective receptive field while controlling computational cost. Third, lightweight ADown modules were used to replace part of the conventional downsampling operations, and the neck network was redesigned as a Cross-Scale Feature Fusion Module with ADown (CCFM_A). This structure was used to enhance multi-scale feature interaction while reducing redundant computation. Experimental results showed that LLDF-YOLO11n achieved a precision of 84.2%, a recall of 84.6%, a mean average precision (mAP) of 88.5% at an intersection over union threshold of 0.5, and an mAP of 60.5% averaged over intersection over union thresholds from 0.5 to 0.95. Compared with the baseline YOLO11n model, recall, mAP at 0.5 intersection over union, and mAP averaged from 0.5 to 0.95 intersection over union increased by 1.8, 1.3, and 1.9 percentage points, respectively. The model contained only 1.40 million parameters and required 4.2 billion floating point operations (FLOPs), representing reductions of 45.7% and 34.4% compared with YOLO11n. After deployment on an NVIDIA Jetson Orin NX edge computing platform using a TensorRT half-precision inference engine, the model achieved an average real-time detection speed of (46.5 ± 1.3) frames per second (FPS) for 640 × 640 video streams. Visual comparison under the three illumination conditions further showed that LLDF-YOLO11n reduced missed detections and false detections relative to the baseline model, particularly in extremely low-light scenes and occluded flower clusters. These results indicate that LLDF-YOLO11n provides a lightweight and real-time visual detection method for yellow dragon fruit flower recognition, supporting subsequent deployment in automated pollination robots under complex field conditions.