Construction of a Wave Image Dataset for Marine Environment Perception

Deep learning currently plays a vital role in wave perception research. However, due to issues such as a scarcity of publicly available datasets, poor generalization capabilities, and weak robustness in existing wave datasets, it remains challenging to train wave perception models with robust generalization abilities. To address this, this study constructed a large-scale image dataset suitable for wave perception in complex environments. It includes a substantial number of real-world infrared wave videos and simulated video images collected from the GX-Encino Waves library. Through preprocessing operations such as video segmentation, frame sampling, image compression, and cropping, and by annotating wave height and period information, a dataset suitable for deep learning model training was formed, named Fusion-Wave. Preliminary training and validation of the dataset using a 3D convolutional neural network demonstrated excellent performance on the test set, exhibiting high accuracy and low error rates. This confirms the Fusion-Wave dataset's strong learnability and its potential to effectively support research on wave parameter perception.