Pedestrian detection in agricultural environments remains an underexplored area in computer vision, particularly in the context of tea plantation management. This study proposes a vision-based personnel detection system customized for tea garden scenarios, aiming to enhance on-site monitoring and labor management. We adopted the YOLOv5 object detection framework and integrated it with the DeepSORT tracking algorithm to develop a real-time pedestrian recognition system. The model was initially trained on the MOT20 dataset and tested on video data collected from real-world tea plantations.Experimental results revealed that models trained on conventional urban pedestrian datasets struggle to maintain detection accuracy in tea garden environments due to occlusions by tea plants, varying lighting conditions, and target sizes. Recognition performance significantly decreased in steep-angle or top-down camera views, highlighting the need for optimized camera deployment and domain-adapted training. To address this, we conducted data augmentation and parameter tuning tailored to the tea plantation context, improving the model’s mean Average Precision (mAP) and tracking consistency. This work contributes a practical framework for customized-area personnel detection in agriculture and underscores the necessity of developing domain-specific datasets for better generalization. The findings provide valuable technical guidance for intelligent agricultural surveillance and pave the way for further research into vision-based monitoring in crop-specific environments.

In precision agriculture systems, camera networks serve as indispensable visual information acquisition devices. However, due to cost constraints and the complexity of field image information, coupled with environmental variability, image quality degradation often occurs, resulting in low-quality visual images. An effective solution for super-resolution (SR) is urgently needed. Traditional image super-resolution methods perform well under laboratory conditions, but their effectiveness is limited in real field environments due to the inability to accurately simulate actual degradation processes. To address this issue, a degradation framework tailored to field scenes is proposed, which involves precise estimation of image degradation features and the design of targeted degradation algorithms. Additionally, drawing inspiration from the powerful global dependency modeling capability of Transformers, an efficient hybrid architecture is designed that combines the local characteristics of convolutional neural networks with the global attention mechanism of Transformers, aiming to improve image reconstruction quality while reducing computational complexity. Experimental results demonstrate that the proposed method outperforms comparative methods, achieving better subjective and objective quality evaluations.

As the birthplace of soybeans, China has become the world's largest importer and consumer of soybeans. The supply and demand imbalance of soybeans has become one of the significant food and oil security concerns in China.To solve this problem, this paper aims to study a soybean germination rate detection method based on image processing. By replacing manual detection of soybean germination rate and measuring soybean root length with machine-based approaches, basic image processing techniques including grayscale conversion, denoising, binarization, dilation, and erosion are applied to preprocess the images of soybean germination. An improved object detection algorithm is used to extract phenotypic features during soybean germination. The main focus of the study is twofold: (1) Measuring the seed area and the minimum bounding rectangle area of the image regions to determine the germination of seeds by comparing their ratio, thus obtaining the germination rate; (2) Using skeleton extraction and embryo-root separation methods to measure the changes in seed root length during germination, thereby selecting seeds with high vitality and improving the quality of soybean seeds.