Wheat Spike Counting Using Regression and Localization Approaches

In-field counting of wheat spike number is a determining indicator in breeder selection and yield estimation. The present article aims to quantify the wheat images via a dotted annotation dataset that can be generated more prompt than popular bounding box and semantic segmentation annotation techniques. Three different hybrid deep learning algorithms are developed by combining the ResNet-34, ResNet-50, and ResNeXt as feature extraction with UNet algorithm as upsampling. The article also examines pretraining the feature extraction on wheat spike counting performance. Additionally, the research investigates regression and localization approaches for counting the wheat spike number using deep learning models. The results indicate a significant improvement in counting performance when pretrained weights are utilized in feature extraction of deep learning models. In addition, ResNeXt-based deep learning model with pretrained weights operates more efficient than other models with MAPE of 1.56% in regression approach and 4.49% in localization approach.