Vector Field Based Algorithm Design for Water Surface Target Spatial Orientation Relation Recognition

Recent research in computer vision has focused on understanding the semantic features of images. To address the spatial orientation relations among water surface targets, which is a fundamental semantic feature, this paper proposes the water surface target spatial orientation vector field (WST-SOVF) algorithm for recognizing spatial orientation relations. The WST-SOVF algorithm encodes these relations into a deep convolutional neural network with two branches: The T-branch employs key-point estimation to classify the target categories, while the S-branch determines the spatial orientation relations between targets by constructing the spatial orientation vector field. A fusion module is also employed to integrates features from both branches to produce a comprehensive triple list of targets, their spatial relations, and confidence levels. Experiments on Huawei’s "Typical Surface/Underwater Target Recognition" dataset demonstrate the algorithm’s excellent performance.