Automatic Recognition of Harmful Algae Images Using Multiple CNN s

Abstract

The monitoring of harmful algae is extremely important for early warning of red tide and protecting water ecological resources. Addressing the problem that manual algae identification is time-consuming, expensive and requires professionals with substantial experience, multiple Convolutional Neural Networks (CNNs) and deep learning based on transfer learning were used to achieve automatic classification of various algae and identification of harmful algae. In this paper, 11 species of harmful algae and 31 species of harmless algae were collected as the input dataset, and transferred to five fine-tuned classical CNN classification models of AlexNet, VGG16, GoogLeNet, ResNet50, and MobileNetV2 for comparison experiments, and finally, the GoogLeN et model reached a relatively higher recognition accuracy. In addition, a new harmful algae identification method was proposed combining the recognition results of five models, and the recall rate is 98.8%. The experiments of this work show that combing multiple CNN s can realize the recognition of harmful algae, which method plays a key role in the preliminary screening of harmful algae.