SwinT-YOLO: Detection of densely distributed maize tassels in remote sensing images

Abstract

The traits and number of maize tassels are essential indicators to monitor the growth status and predict the yield of maize in the practice of agricultural production. The current methods for detecting maize tassels are dependent on Convolutional Neural Networks (CNN) specially developed for extracting features. In view of the failure of CNN to effectively extract global information and accurately detect densely distributed maize tassels in remote sensing images, a new detection model for feature extraction was therefore proposed in this study based on a multi-headed attention mechanism. Data samples were collected from a maize farmland in southwest China using a Unmanned Aerial Vehicle (UAV) along a path scientifically planned according to energy consumption by the UAV, and then processed to label the location and size of each maize tassel in the samples. The labeled data samples were made into a dataset. In response to the disability of CNN to accurately detect densely distributed maize tassels in remote sensing images, SwinT-YOLO was proposed in this study as a new detection model. Swin-Transformer was first used to optimize the backbone feature network of YOLOv4which was followed by the use of the depthwise separable convolution module to reduce the number of parameters and Floating-Point Operations (FLOPs) in the whole detection model. The experimental results show that the new detection model SwinT-YOLO has similar parameters and FLOPs with YOLOv4 and could improve the detection accuracy in the Pascal Visual Object Classes (VOC) dataset by two percentage points. The detection accuracy of SwinT-YOLO also reaches 95.11 % in the homemade remote sensing dataset of maize tassels, suggesting a clear advantage of this model over other excellent detection models. In comprehensive consideration of the detection accuracy, F1-score, and counting results of each model, the new detection model SwinT-YOLO for feature extraction based on a multi-headed attention mechanism is more competent in performing the task of detecting maize tassels.