Accurately and rapidly counting the number of maize tassels is critical for maize breeding, management, and monitoring the growth stage of maize plants. With the advent of high-throughput phenotyping platforms and the availability of large-scale datasets, there is a pressing need to automate this task for genotype and phenotype analysis. Computer vision technology has been increasingly applied in plant science, offering a promising solution for automated monitoring of a large number of plants. However, the current state-of-the-art image algorithms are hindered by hardware limitations, which compromise the balance between algorithmic capacity, running speed, and overall performance, making it difficult to apply them in real-time sensing field environments. Thus, we propose a novel lightweight neural network, named TasselLFANet, with an efficient and powerful structure for accurately and efficiently detecting and counting maize tassels in high spatiotemporal image sequences. Our proposed approach improves the feature-learning ability of TasselLFANet by adopting a cross-stage fusion strategy that balances the variability of different layers. Additionally, TasselLFANet utilizes multiple receptive fields to capture diverse feature representations, and incorporates an innovative visual channel attention module to detect and capture features more flexibly and precisely. We conducted a series of comparative experiments on a new, highly informative dataset called MrMT, which demonstrate that TasselLFANet outperforms the latest batch of lightweight networks in terms of performance, flexibility, and adaptability, achieving an F1 measure value of 94.4%, a mAP.@5 value of 96.8%, and having only 6.0M parameters. Moreover, compared with the regression-based TasselNetV3-Seg† model, our proposed model achieves superior counting performance, with a mean absolute error (MAE) of 1.80, a root mean square error (RMSE) of 2.68, and a R² of 0.99. The proposed model meets the accuracy and speed requirements of the vision system in maize tassel detection. Furthermore, our proposed method is reliable and unaffected by geographical changes, providing essential technical support for computerized counting in the field.

Accurate perception of tea buds is a fundamental prerequisite for intelligent and precise tea harvesting planning. However, in real tea plantation environments, reliable harvesting-oriented perception at the planning level remains highly challenging due to the small size of tea buds, severe occlusion, complex background clutter, and the lack of accurate three-dimensional spatial information. To address these challenges, we propose TeaNeRF, an integrated three-dimensional visual perception pipeline designed for harvesting-oriented tea bud analysis. Instead of treating detection, segmentation, and spatial analysis as independent tasks, TeaNeRF integrates sequential two-dimensional recognition, monocular depth estimation, and neural radiance field reconstruction into a coherent perception pipeline, allowing accurate spatial understanding of tea buds in complex natural scenes. It should be noted that the proposed integration is conducted at the perception-output level, where multiple modular components are connected through fixed interfaces, rather than through joint optimization or an end-to-end trainable formulation. The proposed framework combines an enhanced YOLO-based detector, prompt-guided segmentation, and monocular depth priors to guide NeRF-based three-dimensional reconstruction. By incorporating depth supervision and semantic-aware neural fields, TeaNeRF generates dense and geometrically consistent point clouds with reliable semantic separation. Quantitative evaluations show consistent improvements in reconstruction fidelity, as reflected by increased PSNR and reduced LPIPS across multiple tea tree scenes. Based on the reconstructed semantic point cloud, a three-dimensional clustering and geometric fitting strategy is further developed to enable tea bud counting and harvesting-oriented candidate point estimation at the perception level. Experiments conducted on a real-world dataset of 4,700 tea plantation images demonstrate that TeaNeRF improves detection accuracy (mAP@50 = 91.7%), segmentation quality (IoU = 0.640), and overall three-dimensional perception performance. Case-level counting results on representative tea trees indicate that the proposed 3D semantic point cloud–based approach can provide feasible tea bud counting behavior and consistent spatial guidance cues for downstream harvesting planning. By providing structured three-dimensional spatial information, including tea bud locations, counts, and harvesting-oriented candidate points, TeaNeRF offers practical perception-level outputs for downstream planning in automated tea harvesting systems.

Image deformation refers to deforming objects in images into a target shape or posture. Although point-based image deformation algorithms have made breakthroughs in performance and visual effects, they are limited to warping the current structure information in the source image. For example, while opening a closed mouth in a face deformation, the point-based algorithm cannot generate teeth, which causes the mouth to twist weirdly. Deep learning-based face editing models can generate new parts but cannot achieve fine pixel-level manipulation. To overcome these challenges, we propose a two-step strategy for face deformation. We first generate a high-resolution intermediate image by blending the source image and the specific part of the target image via our face blending generative adversarial network (FB-GAN). Then, we employ a state-of-the-art point-based geometric deformation method to deform the intermediate image with target face guidance. Extensive experiments show that the proposed FB-GAN can generate realistic and high-resolution results and demonstrates that the two-step face deformation strategy can be better applied to human face deformation.

Mainstream loop closure detection based on geometric descriptors struggles to guarantee high detection accuracy in complex environments and requires significant computational resources. To enhance the accuracy of loop closure detection while reducing computational overhead, this paper fully leverages the geometric and dense intensity features obtained from lidar scans to develop a global descriptor that effectively captures the geometric and intensity structures of the environment. An efficient three-stage hierarchical retrieval strategy is employed, including K-Nearest Neighbor search constraints within the ikd-tree and fast geometric and intensity structure reidentification through binary operations. All constraints are globally optimized using factor graphs. The proposed method is evaluated on publicly available datasets, and the results demonstrate that our approach achieves higher trajectory accuracy.

To address the challenges of tea tender bud detection, this paper proposes an improved Real-Time Detection Transformer (RT-DETR) based detection framework. These challenges include the difficulty of extracting discriminative features for small targets, limited spatial modeling under complex natural backgrounds, and inaccurate localization caused by low-quality bounding boxes. Specifically, Cross Stage Partial Darknet (CSPDarknet) is adopted as the backbone network to enhance feature representation capability. Its original C2f modules are restructured by integrating a Dynamic Pyramidal Parallel Attention (DPPA) mechanism into the Bottleneck blocks, resulting in a lightweight C2f-DPPA module. The DPPA incorporates dynamic convolution layers that adaptively aggregate multiple kernels based on input content. This enables flexible and context-aware feature extraction across varying visual conditions.To address insufficient spatial awareness in the attention encoder, a Learnable Positional Encoding (LPE) mechanism is embedded into the AIFI module. This enhances the modeling of positional dependencies under diverse illumination and background patterns. Additionally, an improved Focaler-GIoU loss is introduced to alleviate regression bias in complex scenarios. It places greater emphasis on low-quality samples to enhance bounding box optimization.A self-constructed tea tender bud dataset containing 2500 annotated images was established for training and evaluation. This dataset incorporates diverse conditions including sunny, rainy, foggy, and overcast environments. Experimental results demonstrate that the proposed model, DPPADet-RT, achieves superior detection performance compared to mainstream algorithms on this dataset. It reaches 0.846 Precision, 0.847 mAP@0.5, and 0.366 mAP@0.5:0.95. Moreover, the model maintains a lightweight design with 15.2 million parameters and 34.5 GFLOPs. It achieves an average inference speed of 10 ms per image (100 FPS). These results validate the model’s effectiveness and robustness in real-world agricultural environments.

For scene text recognition (STR), synthetic text image generators have been successfully applied to alleviate the lack of annotated text images from the real world. However, when targeting some challenging text images, their methods have difficulty extracting foreground and background information, and the differences between the synthetic and real samples are large and difficult to adapt. In this paper, we introduce a new end-to-end synthetic text image generator, STIW, by designing a series of synthetic strategies to generate synthetic data that fit realistic text images. In addition, we propose the instantiated weights method, which connects the distributions of different domains by learning and exploiting the instance weights and the Mahalanobis distance metric. This reduces the inter-sample variation in the migration learning process of synthetic text images. Finally, we added an Attention in the LSTM of the recognition algorithm, using a mixed training learning strategy. In our experiments, our method has greater STR boosting performance than the synthetic datasets MJSynth (MJ), SynthText (ST), and SynthTIGER, and the instantiated weight method is able to further improve STR performance. Our ablation study demonstrates the benefits of using individual synthetic strategies, as well as adding to some of the problems of synthetic text image methods.