Underwater Object Detection Using Restructured SSD

Abstract

AbstractDeep learning has been widely used in computer vision tasks such as image classification, semantic segmentation and object detection, which has achieved many breakthrough results in recent years. Compared with conventional object detection tasks, due to objective factors such as uneven illumination, low contrast, and more impurities in the underwater environment, these is no guarantee of high quality for underwater images, which brings challenges to the underwater object detection task. In this paper, we construct an underwater object detection model based on multi-scale feature fusion (called Multi-scale Feature Fusion Network for Underwater Object Detection, MFFNet). MFFNet uses SSD model as the baseline, then makes an improvement by adding three different modules, which are improved FPN, assisting backbone and CBAM attention module. Based on VGG-16 and ResNet-50 as the backbone network, the composite backbone connection is performed; the attention mechanism CBAM module is involved to make the network pay more attention to the objects; the feature pyramid FPN structure is used for multi-scale feature detection. To verify the effectiveness of the network model proposed in this paper, experiments are carried out on three datasets, i.e., VOC 2007, UPRC and Fish4knowledges. The experimental results show that compared with other main object detection models, the network model proposed in this paper has obvious advantages in underwater object detection, and can obtain higher detection accuracy.KeywordsUnderwater object detectionAttention mechanismMulti-scale detection