SAR Image Ship Target Detection Based On Depth Learning

With the development of spaceflight technology,the detection system of Synthetic Aperture Radar(SAR)load in China presents a trend of multi-type and multi-resolution.The synthetic aperture radar with high resolution active sensor can image the sea surface from every angle due to its strong penetrating ability to the atmosphere.So it has become one of the key techniques in the field of image.Since the traditional detection and recognition methods are difficult to adapt to the multi-resolution and multi-type SAR image data,it is necessary to find a method to extract the effective features from the multi-resolution image data.Firstly,aiming at the large error problem of the traditional warship detection method in SAR images,this thesis proposes a reasonable and effective algorithm of edge detection combined with gray difference.The direction of edge is detected by using Canny operator.According to the information of the edge intensity of the Canny operator in the SAR image,the correlation between the detection edge of the ROEWA and the Canny is determined.ROEWA operator is used to detect the edge of the image.As a result,the SAR image is separated from the sea and land by the difference between the sea and the land.Secondly,according to the high application value of convolution neural network model to image detection,the thesis proposes an improved ELU activation function convolution neural network method and establishes a deep learning model that combines the ELU activation function and the quadratic cost function and the distance function between the sample feature and the classification center in the training sample.The experimental results show that the proposed method improves the anti-noise property of ship detection in SAR images,and the detection rate reaches 98.6%.Finally,this thesis compares the disadvantages of Softmax classifier and SVM classifier in SAR images,and then proposes a hybrid fuzzy support vector machine instead of traditional classifier.The effectiveness of the proposed algorithm is obtained by comparing the missing target number,false alarm number and detection efficiency.