Aiming to solve the problem of low recognition accuracy of jujube fruits and poor ability to distinguish between normal and cracked fruits under various complex environmental conditions, such as different lighting, occlusion, weather conditions, and the presence of distant targets, field jujube images in natural environments were collected. Data augmentation was performed by simulating four weather conditions, and a dataset for model training, validation, and testing was created. Based on the YOLO v8n model, the GSConv and RFCAConv modules were incorporated to build the GR-YOLO model, which was then trained and tested on the jujube dataset. The experimental results showed that the GR YOLO model achieved a precision of 95.14%, a recall of 97.71%, and a mean average precision (mAP) of 97.64%. The model size was 4.9 MB, the inference speed was 289.8 f/s, and the number of parameters was 2.38×10^6. Compared with models with the similar number of parameters, including YOLO v5n, YOLO v6n, YOLO v7 tiny, and YOLO v8n, the GR-YOLO model’s precision was improved by 1.82~2.61 percentages, recall was improved by 3.04~8.48 percentages, and mAP was improved by 2.40~7.88 percentages. Compared with existing mainstream models, GR-YOLO model demonstrated optimal recognition performance and a lower number of parameters. The improved model not only increased the recognition accuracy of jujube targets but also effectively distinguished between normal fruits and cracked fruits, while achieving lightweight design.