Study On The Detection Of Bridges For The Large Area Images

In the field of computer vision and machine learning, object detection is an important research area. Military building detection techniques have drawn a wide attention recently, such as airport, port and bridge. These techniques can be used for reconnaissance, military striker and so on. At the same time, geographic system also requires accurate position of these buildings. For an image consisting of more than 4 million pixels and covering more than 40 km wide, it is very hard to detect the bridge, due to the following factors: 1, compared with the entire image, the proportion of the bridge is very small. It is difficult to extract the feature and detect the bridge among the road, mountain, river and other interfering objects. 2, large area images contain too many pixels, traditional methods such as segmentation, edge detection are very time consuming. 3, the bridge itself has many kinds of scale and angel. As a result, how to acquire the accurate position of the bridge in the large area image is a significant and chal enging question.This paper mainly focuses on the bridge detection in the large area images. In order to reduce computational complexity and improve accuracy, the detection is divided into two steps: potential area detection and accurate detection.In the first step, a deep convolutional neural network which has already been trained on image database is used as a “black box” feature description for the bridge sample. Then a classifier is trained on these bridge samples based on the “black box” feature description. In testing procedure, the testing image can be classified by the trained classifier. In the second step, the edge detection technique based on random forests is used to acquire the edge. Then a line extraction method is employed based on the result of the edge detection. At last, the super pixel segmentation based on graph theory and the texture analysis methods are applied to remove false objective and compensate the bridge line.In the end, a kind of method is proposed for the real-time environment. This method has low memory requirement, high processing speed and has already been tested in DSP environment.