Research On Ship Target Detection And Recognition In SAR Image Based On Convolutional Neural Network

As a maritime power,it is very important for China to realize all-day and all-weather ship target detection and identification to safeguard maritime rights and interests.Synthetic Aperture Radar(SAR)has the characteristics of all-day and all-weather operation,which can make up for the deficiency of optical/infrared sensors unable to observe the earth normally at night/rainy day.Therefore,it is of great significance to study target/scene interpretation based on SAR images.Compared with optical image,ship target detection and recognition based on SAR image is more difficult.Based on the characteristics of CNN self-learning target features,this paper systematically studies the methods of ship target detection and recognition in SAR image based on convolutional neural network,and puts forward some meaningful algorithm models.The main research contents of this paper are as follows:In order to normalize the image data during the training of convolutional neural network,a preprocessing method of SAR image based on histogram transformation and improved Lee filter was studied.The image in the detection data set is re-quantized and enhanced by using gray histogram stretching.In order to suppress speckle noise of enhanced SAR image,improved Lee filter is used to denoise SAR image.The YOLO-v3 target detection algorithm is used to test,and the e×periment verifies the effectiveness of preprocessing for ship target detection in SAR image.In order to solve the problems of low recall rate of ships in dense distribution and misdetection in offshore area when detecting ship targets in SAR images,an improved YOLO-v3 ship target detection method was proposed,which integrated convoluted attention module(CBAM)into CSP-Darknet53 and Neck layer respectively.CSPDarknet53 was used to replace Darknet53 of YOLO-v3,and the lower sampling layer of CSP-Darknet53 was fused with CBAM.At the same time,the convolutional layer in the Neck layer and CBAM are combined to replace the original convolutional layer to improve the learning ability of the network for the ship target features.E×perimental results show that the new detection model not only effectively reduces the probability of false detection in the offshore area,but also improves the accuracy of ship target detection in the SAR image with dense distribution.Aiming at the problems of misdetection caused by fuzzy category features and low classification recognition accuracy caused by small samples,a Loss Function of Focal Loss with category weight was proposed.Combined with the improved YOLO-v3 target recognition framework,the target recognition probability can be effectively improved.To satisfy the demand of the SAR images of ship target recognition,the paper constructs a containing five categories of SAR images of ship identification data sets,and use the YOLO-v3,improved network containing CBAM,the improved network with Focalloss+CBAM respectively,which can identify the data set results show that the improved network with Focalloss+CBAM not only improves the false detection of various categories,but also effectively improves the identification precision of categories with fewer samples.