This study addresses the challenge of detecting grain spikes in images for crop yield assessment. Grain spikes are crucial but difficult to discern due to their minimal presence and similarity to surrounding leaves, posing a challenge even for advanced deep-learning networks. We enhance detection by refining the Faster R-CNN (FRCNN) architecture, simplifying feature extraction layers, and adding a global attention module. Testing on diverse European wheat varieties, including complex bushy types, our modified FRCNN (FRCNN-A) outperforms the standard model, achieving a mean Average Precision (mAP) increase from 76.0% to 81.0%. Compared to the Swin Transformer's 83.0% mAP, our model shows competitive performance. We also tested Yolov8 and Yolov4 models, achieving 0.79 and 0.78 mAP, respectively. The model was deployed on the PlantScreenTM system to extract phenotypes such as awn detection and spike count in late reproductive stage plants.