Fig. 2
From: Multi-scale feature fusion keypoint detection network for ship draft line localization
The overall architecture of our proposed method. Initially, our model utilizes two convolutional blocks to extract fundamental features from the input data. The processed data is then fed into four FFEMs. The features produced by FFEMs will undergo multi-scale feature fusion via FWMs. Ultimately, the integrated feature maps are directed into a task-specific head designed for keypoint detection, which predicts the waterline keypoints. These keypoints, combined with character recognition results, are used to compute the final waterline prediction value through a mathematical model.
