The Study On Rough Hough Transform And Neural Network Method For Rectangle Detection

Computer vision is a discipline that perceives information from images or data,one of the most important research content is the detection of graphic.Graphic processing is widely applied in medical image processing,industrial product detection and other fields.For example,when solving complex layout design problems such as satellite cabins,the layout design problem of 3-D cylindrical bodies and cubic objects is first converted into layout design problems of 2-D circular and rectangular objects,and then the human combination of evolutionary algorithm is adopted,the solution to the problem is solved optimally.The method involves the detection of circle and rectangle in layout knowledge graph.In general,firstly the circles on the layout knowledge graph are detected and removed,due to the analytical nature of the circle.rectangles will be incomplete after circles are removed,and false or missed ones will be produced for existing detection methods.In addition,there may be intersecting rectangles on the layout knowledge graph.Therefore,in this dissertation,we studied the problem of rectangle detection of layout knowledge graph under the sponsorship of the Natural Science Foundation of Hunan Province.The main contents are as follows:(1)In this dissertation,a random Hough transform approach is proposed for multi-rectangle detection.First,a pixel is randomly generated as the center of a candidate rectangle,then two pixels is be searched as the diagonal vertices of the rectangle to determine the size of the rectangle,and the third pixel is be searched as the third vertex,according to its position relationship with the two diagonal vertices to determine the inclination of the rectangle,if the two line segments formed by the third vertex and the diagonal vertices are not perpendicular to each other,the vertex will be removed and the invalid sampling can be reduced,and finally determine whether the rectangle formed by three vertices is a real rectangle.The experimental results show that this method is superior to the RTGPIEV algorithm in the detection efficiency and detection accuracy when the image noise points are fewer.(2)In this dissertation,an approach based on hybrid neural network is proposed for multi-rectangle detection algorithm.First,the approximate location of the rectangles is detected by a rectangle detection algorithm based on random Hough transform,and the initial clustering is generated according to the approximate location,then the rotated rectangles are detected based on the single layer neural network and the pixels which belong to the rotated rectangles are removed,then the non-rotated rectangles are detected by a rectangle detection algorithm based on deep learning.The experimental results show that both the detection efficiency and detection accuracy of the proposed method are superior to those of the RTGPIEV algorithm when the image noise points are denser.In this dissertation,the problem of multi-rectangle detection in layout knowledge graph is studied which based on the research of the problem of detecting knowledge graph of satellite cabin layout.A rectangle detection algorithm based on random Hough transform is proposed,it is suitable to the problem of detecting rectangles which are intersecting and incomplete.It is hoped that the rectangle detection method proposed in this dissertation can be extended to more fields and solve the problem of multi-rectangle detection in complex backgrounds in other fields.