<p>In response to the issue of low color contrast against similar backgrounds and occlusion from clustered growth during oblate jujube harvesting, this paper proposes a novel network architecture, termed BEA-Net, which is designed from the ground up for real-time, robust detection in agronomic environments. The proposed method integrated the Bottleneck Transformer (BoT-CTR3), which incorporated multi-head self-attention, enhancing the local feature extraction ability and improving the detection baseline. Subsequently, the hybrid attention mechanism of Efficient Multi-scale Channel Attention (EMCA) was embedded into the backbone network, which improved color feature discrimination under similarly colored backgrounds. Finally, the Adaptively Spatial Feature Fusion (ASFF) module was integrated into the detection head, strengthening the model’s contextual information extraction to compensate for information loss in occluded regions. Experimental results demonstrated that the improved network achieved precision, recall, F1-score, and mean average precision (mAP) of 94%, 92.2%, 93%, and 97.8% respectively—surpassed the baseline network by 3.2, 3.6, 3.0, and 2.1 percentage points. Comprehensive ablation studies validated the contribution of each component, while comparative evaluations against 14 state-of-the-art single-stage detectors confirmed BEA-Net's superior performance. To verify generalizability, the algorithm was tested on our self-built crisp persimmon dataset, achieving 96.9% mAP and 97% F1-score—represented improvements of 2.3 and 4.0 percentage points respectively. Furthermore, deployment on three distinct Android mobile devices achieved an optimal inference time of 50&#xa0;ms. Benchmarking against YOLOv5 variants (n, s, m) confirmed this performance fully satisfies the accuracy and real-time requirements of automated picking systems, significantly enhancing the algorithm's practical application value in real-world scenarios.</p>