Deep-learning Based Ship Detection Method Under Complex Conditions

With a vast sea area of 3 million square kilometers and a coastline of 32,000 kilometers,China is recognized as a big shipping country.The Party and the country value highly the development of marine undertakings and actively promote the construction of intelligent shipping.Ship detection based on visible image,as one of the hot issues in the intelligent shipping field,has important value of application.In recent years,the learningbased generic object detection method has developed rapidly,yet its application to ship detection under complex sea conditions does not achieve favorable results.Moreover,the size of these models is too large to be applied to embedded devices with limited resources.In allusion to the above problems,this paper investigates ship detection based on deep learning,and its main contributions are as follows:(1)In allusion to problems that the current ship detection models based on deep learning endure numerous training parameters and low detection efficiency,this paper puts forward a Light-SDNet based on YOLOv5 for realizing end-to-end ship detection.Light-SDNet,which uses the attention-guided CA-Ghost to extract features and C3 Ghost module to fuse features respectively,can extract abundant ship features via a few operations.The extensive experiments on Sea Ships dataset reveal that compared with the baseline model,the proposed Light-SDNet model possesses less size,lower computation complexity,and better detection results.Compared with other lightweight algorithms,Light-SDNet strikes a balance between model complexity and detection accuracy,which is more suitable for maritime monitoring systems with limited computing power and memory.(2)A hybrid training strategy for generating synthetic degraded images has been proposed to increase the diversity of raw dataset,to enhance the robustness of the model under complex conditions.The proposed strategy enables Light-SDNet to improve ship detection results under severe sea conditions such as fog,rain and low illumination.In addition,compared with other advanced methods on Sea Ships dataset,the proposed Light-SDNet has superior performance in detection accuracy,robustness and efficiency.(3)In allusion to the problem of low detection accuracy of small objects,a multiscale detection network MPH-SDNet is put forward.MPH-SDNet integrates channel features by adopting grouping method and achieves more adequate shallow feature extraction as well as more effective multi-scale feature fusion.Beyond that,the detection heads are rationally configured to detect small-scale ship effectively by using multi-scale features.In order to improve the detection accuracy under complex sea conditions,data combination strategy is utilized to improve the diversity of training data and enhance the generalization of MPH-SDNet.The experimental results show that the proposed MPHSDNet can detect ships efficiently and accurately under complex sea conditions.