A Semi-Self-Supervised Learning Approach for Wheat Head Detection Using Extremely Small Number of Labeled Samples

Abstract

Most of the success of deep learning is owed to supervised learning, where a large-scale annotated dataset is used for model training. However, developing such datasets is challenging. In this paper, we develop a semi-self-supervised learning approach for wheat head detection. The proposed method utilized a few short video clips and only one annotated image from each video clip of wheat fields to simulate a large computationally annotated dataset used for model building. Considering the domain gap be-tween the simulated and real images, we applied two do-main adaptation steps to alleviate the challenge of distributional shift. The resulting model achieved high performance when applied to real unannotated datasets. When fine-tuned on the dataset from the Global Wheat Head Detection Challenge, the performance was further improved. The model achieved a mean average precision of 0.827, where an over-lap of 50% or more between a predicted bounding box and ground truth was considered as a correct prediction. Al-though the utility of the proposed methodology was shown by applying it to wheat head detection, the proposed method is not limited to this application and could be used for other domains, such as detecting different crop types, alleviating the barrier of lack of large-scale annotated datasets in those domains.