Deep learning-based kiwifruit flower recognition method to facilitate automated pollination

Accurate flowering-stage recognition is vital for intelligent orchard management, automated pollination, and yield forecasting. However, in complex natural scenes, existing models struggle to balance detection accuracy with inference speed. To bridge this gap, we propose the Kiwifruit Recognition Network (KiwiRecNet), a lightweight yet high-performance framework tailored to kiwifruit blossoms in the wild. KiwiRecNet first employs the Kiwifruit Generative Adversarial Network for Low-light Improvement (KiwiGAN-LI) to enhance under-exposed images. We then design a novel backbone, the Multi-Scale Shuffle Block (MSBlock), which combines structural re-parameterisation with channel–spatial shuffling to shrink the network footprint. Next, we propose the Partial-Mixing Vision Transformer (PMVIT), a convolution-Transformer hybrid that captures fine-grained features and remains robust to occlusion. Finally, we devise a Bidirectional Cross-Scale Fusion module (Bi-CSF) to enrich multiscale perception. Evaluated on the NWAFU Kiwifruit_F dataset, KiwiRecNet achieves 94.07 % mAP at 82.67 FPS with only 0.93 million parameters, outperforming existing lightweight detectors while approaching heavyweight baselines at a fraction of their cost. Consistent gains across multiple flower datasets confirm its generalisation ability. These results demonstrate an effective route to high-accuracy, real-time flowering-phase recognition on resource-constrained devices, paving the way for scalable agricultural automation. • A low-light image enhancement method, KiwiGAN-LI, was proposed. • A lightweight kiwifruit flowering phase recognition model, KiviRecNet, was developed. • A channel-mixing method, PMVIT, was introduced. • KiviRecNet achieved an optimal speed-accuracy trade-off.