Classification of Aquatic Animals by the Spherical Amphibian Robot based on Transfer Learning

Abstract

The spherical robot is mainly used for normal observation of aquaculture biology. The performance of aquatic biological image recognition mainly depends on the feature extraction and the selected classifier. Traditional manual extraction methods often cannot meet actual application requirements, and have problems such as poor accuracy and weak generalization ability. To solve the above problems, a small data set aquatic animal classification model based on convolutional neural network and transfer learning is proposed in the spherical robot. First, the original images of aquatic animals is preprocessed, and the data set is enhanced using the data increment method. Second, The original CNN model is then improved by embedding the SE module and using the triplet loss function to replace the softmax loss function. Finally, Transfer learning a deep pre-trained model of the ImageNet image data set. Training and fitting parameter distributions on aquatic image data sets. Experimental results show that the model optimizes the accuracy of aquatic animal target recognition, and the test accuracy reaches 93.11%.The model has good stability and high precision in aquaculture environment.