Concatenate and Shuffle Network: A Real-Time Underwater Object Detector for Small and Dense Objects

Object detection of underwater optical images is of great significance in many underwater missions, such as the salvage of underwater objects, the exploration of marine organisms, etc. However, underwater objects are often small and dense, which are difficult to detect. To tackle above issues, we propose a novel framework of underwater object detection named Concatenate and Shuffle Network (CSNet) based on center points detection, which can not only detect small and dense objects with high accuracy, but also detect in real time. Firstly, a multi-scale fusion strategy called Feature Concatenation Shuffle (FCS) is proposed. The detailed features from shallow layer in Convolutional Neural Network are completely integrated into deep layer, and the capability for extracting features of small objects is enhanced. Moreover, to accelerate our method, we propose a lightweight deconvolution block (DB), which integrates a structure of dual-branch feature fusion and a lightweight deconvolution method. In addition, we study the advantages of detecting dense objects based on center points and introduce it to our detector. Lastly, experiments show that CSNet achieves the best speed-accuracy trade-off on URPC 2018 with 39.7% AP at 58.8 FPS and 42.4% AP with multi-scale testing at 5.7 FPS. Compared with several state-of-the-art detectors, CSNet reaches a competitive accuracy at a breakthrough speed and can run in real time under various computing conditions.