We propose a deep learning tool to localize fish objects in benthic underwater videos on a frame by frame basis. The deep network predicts fish object spatial coordinates and simultaneously segments the corresponding pixels of each fish object. The network follows a state of the art inception resnet v2 architecture that automatically generates informative features for object localization and mask segmentation tasks. Predicted masks are passed to dense Conditional Random Field (CRF) post-processing for contour and shape refinement. Unlike prior methods that rely on motion information to segment fish objects, our proposed method only requires RGB video frames to predict both box coordinates and object pixel masks. Independence from motion information makes our proposed model more robust to camera movements or jitters, and makes it more applicable to process underwater videos taken from unmanned water vehicles. We test the model in actual benthic underwater video frames taken from ten different sites. The proposed tool can segment fish objects despite wide camera movements, blurred underwater resolutions, and is robust to a wide variety of environments and fish species shapes.
展开▼
