Fast Accurate Fish Detection and Recognition of Underwater Images with Fast R-CNN

This paper aims at detecting and recognizing fish species from underwater images by means of Fast R-CNN (Regions with Convolutional Neural and Networks) features. Encouraged by powerful recognition results achieved by Convolutional Neural Networks (CNNs) on generic VOC and ImageNet dataset, we apply this popular deep ConvNets to domain-specific underwater environment which is more complicated than overland situation, using a new dataset of 24277 ImageCLEF fish images belonging to 12 classes. The experimental results demonstrate the promising performance of our networks. Fast R-CNN improves mean average precision (mAP) by 11.2% relative to Deformable Parts Model (DPM) baseline-achieving a mAP of 81.4%, and detects 80x faster than previous R-CNN on a single fish image.