G-Net: An Efficient Convolutional Network for Underwater Object Detection

Abstract

Visual perception technology is of great significance for underwater robots to carry out seabed investigation and mariculture activities. Due to the complex underwater environment, it is often necessary to enhance the underwater image when detecting underwater targets by optical sensors. Most of the traditional methods involve image enhancement and then target detection. However, this method greatly increases the timeliness in practical application. To solve this problem, we propose a feature-enhanced target detection network, Global-Net (G-Net), which combines underwater image enhancement with target detection. Different from the traditional method of reconstructing enhanced images for target detection, G-Net realizes the integration of image enhancement and target detection. In addition, our feature map learning module (FML) can effectively extract defogging features. The test results in a real underwater environment show that G-Net improves the detection accuracy of underwater targets by about 5%, but also has high detection efficiency, which ensures the reliability of underwater robots in seabed investigation and aquaculture activities.