Research On Ship Recognition Based On Deep Learning

With the continuous development of the maritime industry,maritime traffic is becoming more and more frequent.It is particularly important to do a good job in maritime security、marine environment monitoring、marine vessel monitoring and related research.Nowadays,domestic and foreign countries mainly identify ships based on SAR(Synthetic Aperture Radar)images.It is mainly based on traditional machine learning methods,but with the rapid development of deep learning in recent years,more and more scholars have begun to use deep learning methods to carry out research on SAR image ship recognition work.Due to the high resolution of optical images and the advantages of better visual judgment and more feature information,the ship image recognition has also become a new research direction of ship identification,so it is based on deep learning methods,combined with SAR images and optical images of ships to carry out research on ship monitoring and classification is of great significance.This article focuses on using deep learning methods to do research work on marine ship recognition.The main contents are as follows:(1)A ship recognition method based on improved region fully convolutional neural network(R-FCN)was proposed,which was applied to the optical image ship data set.Firstly,according to the characteristics of small target and a great number of ship recognition targets,The number of the feature layer was increased to 21,which improved the dimension of feature information and avoided the possibility of model imbalance;Secondly,we changed the global average pooling in the ROI Pooling layer to global maximum pooling,which increased the amount of feature information and effectively retained the image information features of small targets and avoided a part of the mean shift caused by errors in convolutional layer parameters;Finally,we added the ROIAlign feature and proposed to cancel the quantization operation,we used bilinear interpolation to calculate feature points to transform the feature extraction process into a continuous feature sampling process,which greatly improved the performance of small target ship detection.This method effectively improved the recognition rate of ships.(2)Based on the YOLO v3 network,a SAR image ship recognition method based on depth feature enhancement was proposed,and the performance of SAR image ship detection and recognition before and after enhancement was compared.Firstly,the feature extraction depth was strengthened,and the feature information of the target image was enhanced by modifying the position of the upsample layer and increasing the number of convolution layers.Secondly,the pre-anchor frame anchor was recalculated for the existing image data set to further improve the accuracy and enhance the position information of the target image.Finally,in the regression frame,the GIoU method was used to replace the traditional IoU regression,which effectively improved the accuracy of the target regression and enhanced the regression information of the target image.Compared with other detection methods,the performance of mAP and FPS were greatly improved.(3)An improved SAR image ship recognition method based on feature information was proposed,and the performance of SAR image ship detection and recognition before and after feature information improvement was compared.Firstly,we introduced Xception and EfficientNet B7 to adjust the network structure,so that the size of the imported network conformed to the original YOLO v3 network structure,and then replaced the original DarkNet-53 network,thus strengthening the effectiveness of feature extraction.Secondly,we used ImageNet to remove the network weights of the fully connected layer to stabilize the situation of large fluctuations of Loss in the initial training period,thereby enhancing the stability of feature extraction.Finally,the DIoU algorithm was introduced to replace the original GIoU algorithm,which further improved the efficiency of the regression frame of the network,thereby enhancing the accuracy of the regression frame of small target ships.This method maximized the performance of the model in the SAR image data set.