A review of weed image identification based on deep few-shot learning

Weeds plague the growth and yield of crops, which is a major obstacle to agricultural production. In recent years, deep learning although has made great breakthroughs on weed image identification, it still faces notable limitations in dealing with complex farmland environments, morphological diversity of weeds, and limited samples, such as data scarcity, few labeled data, and many weed categories. The important reason is that deep learning models need to rely on huge datasets for training, which makes the work of weed data collection huge. Few-shot learning of weed images can use deep learning model to learn effective patterns from few-shot and different classes of weeds, so as to solve the problem of poor performance of depth model in weed image identification under the condition of limited samples. From the perspective of weed identification by intelligent agricultural machinery, this paper tracks the research progress of few-shot learning in recent years and discusses the methods and strategies of using deep learning technology to achieve efficient and accurate weed identification under few-shot. It focuses on the corresponding strategies of few-shot learning models in datasets, feature extraction, and image classification, including data augmentation, meta-learning, active learning, metric learning, and transfer learning. Finally, this paper analyzed the application cases of weed recognition based on deep few-shot learning in real farmland scenarios in recent years and explored the performance, advantages, and disadvantages of few-shot learning methods in the application of vehicle and airborne platforms. This paper finds that the Siamese network can learn features using only 10% of the data through comparative analysis. Data augmentation and active learning effectively address class imbalance, but meta learning excels in predicting unseen classes. Fine-tuning is widely used and performs best in 5-shot scenarios. This research not only fills the research gap of few-shot learning of weed recognition but also provides new ideas and methods based on limited data. Looking forward to the future, we hope to provide useful reference for the research of weed identification driven by few-shot data.