Multilayer Feature Extraction Network for Military Ship Detection From High-Resolution Optical Remote Sensing Images

Abstract

Rapid and accurate detection of maritime military targets is of great significance for maintaining national defense security. Few studies have used high-resolution optical images for the detailed classification of maritime military targets. This article, inspired by EfficientDet trackers, presents a method to classify military targets on the sea from high-resolution optical remote sensing images. In the first stage, a multilayer feature extraction network is constructed to extract various features. At the same time, residual connection and dilation convolution are introduced to prevent the deep network features from disappearing. Moreover, we use multilevel attention mechanism approaches to make more effective use of multilayer features. ReLU is introduced to replace the original swish activation function to reduce the computational cost in the pretreatment stage. After this, deep feature fusion networks and prediction networks are constructed to locate and distinguish different types of ships. Different types of ships use different degrees of data expansion methods to solve the problem of sample shortage and imbalance. The multiclassification method is used to solve low classification accuracy caused by little difference between civil and military ships. Experimental results suggested that the proposed method can accurately identify multiple types of military ships.