Deep but lightweight neural networks for fish detection

The explosive growth of the underwater images make the demand for automatic accurate object detection more and more urgent. In this paper, we introduced a deep but lightweight neural network to detect fishes. It achieved the state-of-the-art accuracy for fish detection on the dataset of ImageCLEF, which includes 24,277 fish images belonging to 12 classes. Compared with the common used detection network, such as Faster R-CNN, we change the structure of convolution layers by using some building blocks including concatenated ReLU, Inception, and HyperNet. The final network obtained best results of 89.95% mAP(mean average precision), 7.25% higher than the Faster R-CNN network on the same dataset.