Vision based real-time fish detection using convolutional neural network

Abstract

Underwater vision has specific characteristics such as high attenuation of lights, severe noise and haze in the images. For real-time fish detection using underwater vision, this paper proposes convolutional neural network based techniques based on You Only Look Once algorithm. Actual fish video images were used to evaluate the reliability and accuracy of the proposed method. As a result, the network recorded 93% classification accuracy, 0.634 intersection over union between predicted bounding box and ground truth, and 16.7 frames per second of fish detection. It also outperforms another fish detector using sliding window algorithm and classifier trained with histogram of oriented gradient features and support vector machine.