WheatSpikeNet: An Improved Wheat Spike Segmentation Model for Accurate Counting from Field Imaging

This study proposes a meticulously curated and annotated dataset, taken from the publicly accessible SPIKE dataset, and an optimal instance segmentation approach named as WheatSpikeNet for segmenting and counting wheat spikes from field imagery. Proposed method is based on the well-known Cascade Mask RCNN architecture with model enhancements and hyperparameter tuning to provide state-of-the-art detection and segmentation performance. A comprehensive ablation analysis incorporating many architectural components of the model was performed to determine the most efficient version. In addition, the model's hyperparameters were fine-tuned by conducting several empirical tests. ResNet50 with Deformable Convolution Network (DCN) as the backbone architecture for feature extraction, Generic RoI Extractor (GRoIE) for RoI pooling, and Side Aware Boundary Localization (SABL) for wheat spike localization comprise the final instance segmentation model.  With bbox and mask mAP scores of 0.93 and 0.9404, respectively, on the test set, the proposed model achieved superior performance on the challenging SPIKE datasetset. Furthermore, in comparison with other existing state-of-the-art methods, proposed model achieved up to a 0.38% improvement in spike detection and a 3.6% improvement in the segmentation tasks.