Abstract: The accurate detection of kiwifruit flower buds and blossoms across the full flowering period was critical for the autonomous operation of intelligent bud thinning and pollination robots. However, in complex agricultural scenarios, the continuous morphological evolution and periodic overlap of kiwifruit floral organs throughout the full flowering period posed significant challenges to the generalization capability of existing detection models. Specifically, the gradual transition from tightly closed buds to fully open blossoms introduced substantial intra-class variation and inter-class ambiguity, while the asynchronous developmental stages of individual flowers within the same orchard further exacerbated the difficulty of maintaining consistent detection performance. To address these limitations, this study proposed a generalized detection technical approach for kiwifruit bud-flower targets that leveraged full-phase continuous image data and fused phenological temporal information.An automated image acquisition and management system for the full flowering period was designed and deployed to perform automatic data collection and systematic management of kiwifruit floral organs. This system enabled the construction of a full-phase kiwifruit bud and flower image sequence dataset with strict spatiotemporal alignment characteristics, ensuring that each image sequence captured the phenological progression of identical floral targets under consistent environmental conditions. Based on the Swin-T-D (Swin-Transformer-Detection) model, three key technical innovations were introduced. First, a Spatial-to-Depth (SPD) convolution downsampling module was incorporated to replace the patch merging layer of the standard Swin Transformer, which enhanced feature retention capability for small bud targets by preserving fine-grained spatial information during the downsampling process. Second, a Temporal-Aware Encoding (TAPE) mechanism was designed to establish a phenological development manifold, which encoded the temporal continuity of morphological changes and reduced misdetections caused by phenological growth variations of buds and blossoms. Third, a Weighted Adaptive Feature Fusion (WAFF) strategy was proposed to optimize multi-scale feature fusion for bud and flower targets through dynamic weight allocation, thereby adaptively emphasizing discriminative features across different scales and developmental stages. These three modules were integrated to form the Swin-T-kiwiflower detection model, which incorporated phenological temporal cues into the detection framework.Comprehensive comparative experiments were conducted to evaluate the performance of the proposed Swin-T-kiwiflower model against several state-of-the-art detection architectures, including YOLOv8n (You Only Look Once version 8 nano), Mask Region-based Convolutional Neural Network (Mask R-CNN), and the baseline Swin-T-D model. The experimental results demonstrated that Swin-T-kiwiflower achieved superior detection performance across all evaluated metrics. Specifically, the Average Precision (AP) for flower bud detection reached 84.9%, representing a substantial improvement over the baseline Swin-T-D model and outperforming both YOLOv8n and Mask R-CNN by considerable margins. The AP for blossom detection attained 93.4%, indicating robust recognition capability for fully developed floral targets. The overall mean Average Precision (mAP) at an Intersection over Union (IoU) threshold of 0.5 reached 89.2%, which validated the effectiveness of the integrated optimization framework. In terms of computational efficiency, the single-frame inference latency was controlled within 67 milliseconds (ms), demonstrating that the proposed model maintained real-time processing capability despite the incorporation of temporal modeling components. Ablation studies were further performed to quantify the individual contributions of the three proposed modules. The results indicated that the SPD module significantly improved the detection accuracy for small bud targets by mitigating information loss during feature downsampling. The TAPE mechanism effectively reduced false positives and false negatives arising from phenological morphological transitions by leveraging temporal consistency constraints. The WAFF strategy enhanced the adaptability of multi-scale feature aggregation, particularly for targets exhibiting large scale variations between tightly closed buds and expanded blossoms. The combined SPD-TAPE-WAFF triple optimization mechanism demonstrated strong synergistic effects, with the full Swin-T-kiwiflower model outperforming all ablated variants.The proposed Swin-T-kiwiflower model successfully addressed three fundamental challenges in kiwifruit bud-flower detection: the retention of fine-grained features for small bud targets, the adaptation to phenological temporal evolution, and the accommodation of scale and morphological variations between buds and blossoms. The validation results confirmed the effectiveness of the SPD-TAPE-WAFF triple optimization mechanism in enhancing detection generalization across the full flowering period. With a bud AP of 84.9%, a blossom AP of 93.4%, an overall mAP of 89.2%, and a per-frame inference latency of 67 ms, the model achieved an optimal balance between detection accuracy and computational efficiency. This study provided a robust technical foundation for the reliable operation of intelligent bud thinning and pollination robots throughout the full flowering period, and the proposed methodological framework offered a transferable paradigm for phenology-aware object detection in other perennial fruit crops.