Abstract: Yellow dragon fruit (Selenicereus megalanthus) is one of the highly prized cactus vines in tropical areas. Its flowers open at night, which makes detection for pollination difficult under weak illumination, dense flower distribution, and branch occlusion. This study aims to propose 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. The model also supports real-time visual perception for pollination robots in the field. An image dataset was acquired from multiple scenarios at a commercial yellow dragon fruit planting base in Saima Dam, Lujiang Town, Baoshan City, Yunnan Province, China. A total of 1,988 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 flowering stages relevant to pollination. All images were center-cropped to a 1:1 aspect ratio and then resized to 720×720 pixels to standardize image scale and reduce background interference. An illumination labelling strategy was adopted to improve annotation reliability under various scenarios. Images collected under supplemental array lighting and natural light were annotated directly with minimum bounding boxes. Furthermore, brightness-assisted labelling was used for extremely low-light images. The original images were first processed with brightness and exposure compensation for more visible target contours; the label coordinates were mapped back to the original low-light images after manual annotation. A total of 1,600 images were retained as the basic dataset after removing blurred and highly duplicated samples. The dataset included 524 supplemental-light, 530 extremely low-light, and 546 natural-light images. After that, 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, 960 validation, and 480 test images. The LLDF-YOLO11n model was designed for 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. (1) A Self-Calibrated Illumination (SCI) module was inserted into layer 0 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. The detection network benefited from the illumination-corrected input images, thereby preserving the enhancement effect of the pretrained module. (2) The original attention structure was replaced with a C2PSA_LSKA module, in which Large Separable Kernel Attention (LSKA) was introduced to simulate long-range spatial dependencies. Large convolution kernels were decomposed into separable one-dimensional and dilated depthwise convolutions. Thereby, the effective receptive field was enlarged while computational costs were reduced. (3) Lightweight ADown modules were used to replace the conventional downsampling. The neck network was redesigned as a Cross-Scale Feature Fusion Module with ADown (CCFM_A). Multi-scale feature interaction was enhanced to reduce redundant computation. Experimental results showed that LLDF-YOLO11n achieved a precision of 84.2%, a recall of 84.6%, and mAP_0.5 and mAP_0.5-0.95 of 88.5% and 60.5%, respectively. Compared with the baseline YOLO11n model, recall, mAP_0.5, and mAP_0.5-0.95 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 frame rate of 46.5 FPS for 640 × 640 video streams. Visual comparison under the three illumination conditions further showed that LLDF-YOLO11n reduced missed 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.