| At present,our country’s river basins are seriously polluted,and the water surface is flooded with garbage.The situation is severe,which has a huge impact on the city appearance and water quality.With the current construction of smart cities in full swing,unmanned cleaning boats,as an important part of smart water services,have gradually replaced manual cleaning.It has become a hot topic to solve the problems of image segmentation and floating object detection in the process of unmanned ship.In this paper,a floating object detection system is designed for this problem.In this paper,a data set of waterfront segmentation and floating object detection is produced.The samples are mainly collected on-site during the work of the unmanned ship,and the data set is enhanced.Aiming at the characteristics of waterfront image segmentation that is greatly affected by color features,a generative confrontation network is used to generate winter waterfront segmentation samples,and small sample expansion is carried out for the characteristics of small samples in the detection of floating objects.Aiming at the problem of waterfront image segmentation,two semantic segmentation algorithms,PSPNet and DeepLabV3+,are used for training and testing,and the better DeepLabV3+ algorithm is selected.Aiming at the slow segmentation speed and poor segmentation effect of DeepLabV3+ algorithm,an improved DeepLabV3+ algorithm is proposed,which uses lightweight backbone network Mobile Net V3 and improved ASPP to improve segmentation speed and accuracy.Aiming at the problem of YOLOv4’s poor detection effect on small targets in the data set,a feature fusion structure fusing low-level semantic information is proposed.The experiment proved it that the detection of small targets in the data set has a higher average accuracy and a better effect.The design and construction of each module of the floating object detection system have been completed,and the feasibility of the system has been verified. |
PDF Full Text Request