Multi-Scale Feature Fusion Model for Wheat Spike Counting

Predicting wheat yield based on wheat spike images is highly significant for breeding, cultivation, and agricultural production. Counting the number accurately of wheat spikes per unit area is a crucial component in monitoring yield. A novel approach utilizing a deep convolutional neural network with multi-scale feature fusion for the automatic localization and regression of wheat spike counting is proposed. Initially, the point labeling-based P2Pnet (Point-to-Point Network) is employed as the backbone to construct the basic counting framework. Furthermore, considering the variation in the size of wheat spikes, multi-scale information is preserved by extracting features from different depth feature layers of the VGG16 network. To tackle the challenge of tiny wheat spikes, a Triplet Attention mechanism is added to the multi-scale feature fusion module, facilitating the interaction of information across the channel, height, and width dimensions to prevent the loss of features from tiny objects during the learning process. Experimental results show that on the public dataset, the Mean Absolute Error (MAE) achieved by the proposed method is 13.827, which is 20.6% lower than that of the basic P2Pnet. Compared with the single-stage object detection algorithm YOLOV5, the MAE is reduced by 9.28%. Compared with the two-stage object detection algorithm Faster R-CNN, the MAE is reduced by 1.24%.