| The intelligence degree of fishing harbor is an important criterion to measure the modernization level of coastal cities.With the rapid development of emerging technologies such as artificial intelligence,5G,radar,etc.,the fishing harbor intelligent supervision system can realize the rapid dynamic detection and intelligent recognition of fishing ships based on the integration of monitoring video data and AIS data.It fundamentally improves the operational efficiency of fishing harbor and ensure the operational safety of fishing ships.However,there are some problems in the traditional fishing harbor supervision system,such as messy distribution of multi-channel video monitoring screens,easy to overlap or blind areas,and difficult to observe comprehendsively.Moreover,the fishing ship detection and recognition equipment is expensive,not easy to maintain,and the communication signal is easy to be interfered.In view of the deficiencies and problems of the traditional fishing harbor supervision system,this topic is oriented to the intelligent construction of fishing harbor,in which the key technologies based on visual perception and recognition are studied.The main research contents include:1.Proposed a panoramic stitching algorithm for large parallax scenes.Large parallax image stitching often suffers from the problem of difficult alignment of different depth layers.To this end,a hybrid warping model with layered warping and structure preservation is proposed to optimize alignment and preserve local and global structure.To further alleviate the artifacts caused by local misalignment of parallax images,an optimal suture search algorithm is proposed.Finally,quantitative and qualitative analysis of the algorithm performance is done on public and autonomously collected datasets.And the algorithm is applied to realize panoramic stitching of different harbor scenes,both with satisfactory results.2.Proposed a YOLOv3-based object detection method for fishing ships with panoramic images.A panoramic image fishing ship object detection method is developed to address the problem that the same fishing ship appears in multiple monitoring screens at the same time,which leading to repeated detection and difficult statistics.This method adds a multi-scale detection network to YOLOv3 network,which makes it more suitable for small target detection under panoramic images.At the same time,a new K-means clustering algorithm is introduced to improve the detection accuracy and efficiency.Finally,the algorithm before and after the improvement is compared on the panoramic fishing ships dataset,and the improved algorithm achieved better results.3.The character recognition algorithm based on DBNet character detection algorithm and CRNN is studied.Considering the existence of angular tilt,variable position and various scales of fishing ship license plates,etc.This thesis the character detection algorithm DBNet is applied to locate the character area of ship license plates,and the robustness of the algorithm in ship license plate characters detection is verified.To solve the problems of different degrees of blurring and defacement of ships plate characters,a character recognition network based on CRNN was proposed to improve the recognition accuracy of the network.Finally,the effectiveness of DBNet algorithm and the improved CRNN algorithm are verified on the self-built ship license plate dataset.The experimental results show that the character recognition accuracy is improved by2.67% compared with the original algorithm,and it has better robustness in fishing ships license plate character recognition. |
PDF Full Text Request