Tropical Coral Reef Fish Identification Method Based on Resnet Unconstrained Environment of the Seabed

  • Shan Zhang
Keywords: Resnet, Deep Learning, Unconstrained Environment, Coral Reef Fish, Image Recognition


Submarine unconstrained environment video is shot in uncontrolled open seas. There are multi-mode
backgrounds, complicated lighting and weather changes, and rapid growth of algae attached to the lens, which
affects the stability of video quality, resulting in difficulty in image recognition. The performance of algorithms
is generally better than other methods, and it is necessary to build models in combination with specific scenarios
and applications. This paper proposes a tropical coral reef fish identification method based on ResNet deep
learning. Firstly, three indexes of brightness, smoothness and hue are calculated, and the image is preprocessed
with defogging algorithm to improve the quality. Secondly, deep convolutional neural network is constructed
based on ResNet to study the influence of different learning efficiency and network depth on the recognition
results. The results show that the number of neural network layers increases, the training loss and the
verification set loss function will decrease accordingly, but the effect of learning efficiency on the final result is
much weaker; the learning efficiency decreases, the training loss function converges obviously, and the proof set
loss function fluctuates around 10%, which can better accomplish the identification of tropical coral reef fish in
the unconstrained environment of the seabed.