Research On Automatic Ship Detection Based On Improved Grabcut From UAV Images

As UAVs are widely used in ship positioning and regulatory fishery activities,aerial image data quantity is increasing. How to detect and locate ships efficiently andaccurately from the aerial image data has become urgent needs of industrialprocessing. In this paper, some key problems are researched and some results areobtained.For sea template selection problem, a sea template automatically selected modelis proposed. In this paper, images of different area, climate conditions andillumination condition are selected to build surface template library. Images intemplate library not only represent the basic property of the sea, but also determinethe characteristics of the surface of the background and foreground threshold.An approach to the sea background model construction problem is proposed.Region growth algorithm is used to automatically generate background Trimap. Basedon the template detected, SSQ algorithm is used to find the seed point informationautomatically, which solves the problem that seeds artificial provided arerandomness, uncertain, and weak descriptive to sea surface.To solve the issue of accessing sea background automatically, Grabcut algorithmbased on background model is proposed. Background Trimap is used to initializeGrabcut algorithm, which meets the need to process large data in batch and solves theproblem that initial rectangle of Grabcut is manually provided. While for thesegmentation process, repeatedly iterative processes are no longer necessary inaccurate separation of background and foreground.For the candidate domain classification problem, the ships candidates’recognition model is presented. Images after segmentation with improved Grabcut areclassified according to shape of candidate domain and pixel domain size to filter thesimple-connect candidate to achieve the identification of ships.In summary, this paper proposes a new ship detection method. Comparing withthe recognition method that building training library to directly classify has greaterimprovement in precision, comparing with the direct use of image segmentationalgorithm there are some improvement in accuracy, it can be applied to detect ships inindustrial aerial image.