In rice cultivation and breeding, obtaining accurate information on the quantity and spatial distribution of rice plants is crucial. However, traditional field sampling methods can only provide rough estimates of the plant count and fail to capture precise plant locations. To address these problems, this paper proposes P2PNet-EFF for the counting and localization of rice plants. Firstly, through the introduction of the enhanced feature fusion (EFF), the model improves its ability to integrate deep semantic information while preserving shallow spatial details. This allows the model to holistically analyze the morphology of plants rather than focusing solely on their central points, substantially reducing errors caused by leaf overlap. Secondly, by integrating efficient multi-scale attention (EMA) into the backbone, the model enhances its feature extraction capabilities and suppresses interference from similar backgrounds. Finally, to evaluate the effectiveness of the P2PNet-EFF method, we introduce the URCAL dataset for rice counting and localization, gathered using UAV. This dataset consists of 365 high-resolution images and 173,352 point annotations. Experimental results on the URCAL demonstrate that the proposed method achieves a 34.87% reduction in MAE and a 28.19% reduction in RMSE compared to the original P2PNet while increasing $R^2$ by 3.03%. Furthermore, we conducted extensive experiments on three frequently used plant counting datasets. The results demonstrate the excellent performance of the proposed method.

Wheat breeding heavily relies on the observation of various traits during the wheat growth process. Among all traits, wheat head density stands out as a particularly crucial characteristic. Despite the realization of high-throughput phenotypic data collection for wheat, the development of efficient and robust models for extracting traits from raw data remains a significant challenge. Numerous fully supervised target detection algorithms have been employed to address the wheat head detection problem. However, constrained by the exorbitant cost of dataset creation, especially the manual annotation cost, fully supervised target detection algorithms struggle to unleash their full potential. Semi-supervised training methods can leverage unlabeled data to enhance model performance, addressing the issue of insufficient labeled data. This paper introduces a one-stage anchor-based semi-supervised wheat head detector, named “Wheat Teacher”, which combines two semi-supervised methods, pseudo-labeling, and consistency regularization. Furthermore, two novel dynamic threshold components, Pseudo-label Dynamic Allocator and Loss Dynamic Threshold, are designed specifically for wheat head detection scenarios to allocate pseudo-labels and filter losses. We conducted detailed experiments on the largest wheat head public dataset, GWHD2021. Compared with various types of detectors, Wheat Teacher achieved a mAP0.5 of 92.8% with only 20% labeled data. This result surpassed the test outcomes of two fully supervised object detection models trained with 100% labeled data, and the difference with the other two fully supervised models trained with 100% labeled data was within 1%. Moreover, Wheat Teacher exhibits improvements of 2.1%, 3.6%, 5.1%, 37.7%, and 25.8% in mAP0.5 under different labeled data usage ratios of 20%, 10%, 5%, 2%, and 1%, respectively, validating the effectiveness of our semi-supervised approach. These experiments demonstrate the significant potential of Wheat Teacher in wheat head detection.

Pruning and compression of models are practical approaches for deploying and applying deep convolutional neural networks in scenarios with limited memory and computational resources. To mitigate the impact of pruning on model accuracy and enhance the stability of pruning (defined as the negligible drop in test accuracy immediately following pruning), an algorithm for reward-penalty decoupling is introduced in this study to achieve automated sparse training and channel pruning. During sparse training, the influence of unimportant channels is automatically identified and reduced,thereby preserving the ability of the important channels for feature recognition. First, by utilizing the gradient information learned through network backpropagation, the feature scaling factors of the batch normalization layers are combined with the gradient to determine the importance threshold for the network channels. Subsequently, a two-stage sparse training algorithm is proposed based on the reward-penalty decoupling strategy, applying different loss function strategies to the feature scaling factors of "important" and "unimportant" channels during decoupled sparse training. This approach has been experimentally validated across various tasks, baselines, and datasets, demonstrating its superiority over the previous state-of-the-art methods. The results indicate that the effect of pruning on model accuracy is significantly alleviated by the proposed method, and pruned models require only limited fine-tuning to achieve excellent performance.