Research On Detection And Classification Of Ships In Remote Sensing Images Based On Deep Learning

Optical remote sensing images have high resolution,strong timeliness,rich background information,many types of objects and explosive growth of data volume,which have important research value in the civilian and military fields.Ship detection based on traditional methods has the problems of redundant candidate regions,poor robustness of artificially designed features,low accuracy,poor scene applicability and low intelligence.There are many kinds of target detection algorithms based on deep learning,and they have been widely used in the field of image recognition and have achieved great success.This paper attempts to apply deep learning technology to the field of optical remote sensing image ship detection,and mainly carries out the following three aspects:First,the construction of the dataset.There are few datasets dedicated to ship detection publicly available on the Internet,and remote sensing images of military ships are relatively sensitive and difficult to obtain.We have collected 7,072 images from mainstream optical remote sensing image datasets,Google satellite remote sensing images and the Internet.We performed 5-fold data augmentation on 510 remote sensing images containing military ships using the strategy of flipping,rotating and scaling,and then unified the size,format and naming of the images,and finally completed the target labeling of 9622 images.Some images of the dataset have noise interference,the ship target is close to the background information of the image,the image is dark and the cloud is occluded.This paper gives the corresponding solution,and the feasibility of the image preprocessing scheme is verified in the fourth chapter.Second,research on ship detection in different scenarios.There are many kinds of target detection algorithms based on deep learning.In the single-category target research stage,Faster R-CNN,YOLO v3,YOLO v4,YOLO x and SSD algorithms were selected,and the feature extraction network of Faster R-CNN was selected by VGG-16 replaced with ResNet-50 for comparison.We used a single-class dataset to study the performance of different algorithms on the single-vessel pure sea background,multi-vessel pure sea background,single-vessel landing background and multi-vessel landing background test sets.The experimental results show that:in the overall test of the test set,the improved Faster R-CNN has the highest recall rate of 98.44%,and the improved Faster R-CNN has the lowest accuracy rate of 64.85%,which is far lower than the YOLO series and SSD algorithms,indicating that the Faster R-CNN has the strongest ship detection ability,but there are a lot of false detections.SSD has the lowest recall rate of 89.35%,which is because SSD has poor detection effect on small and medium sized ships,and there are many missed detection phenomena.The average precision,recall rate and detection speed of the YOLO series are not much different,and the precision rate of YOLO x is 4.17%and 3.75%lower than that of YOLO v3 and YOLO v4,respectively.Finally,we choose YOLO v3 and YOLO v4 for the classification and identification of ships.Third,research and improvement of classification and identification of ships.In the multi-category target research stage,YOLO v3,YOLO v4 and YOLO v4-tiny are used for classification and identification of ships.In order to improve the overall performance of the algorithm and make it more suitable for small ship target detection,a hybrid attention mechanism is added after the feature output layer of the original network,so that it pays more attention to the details of the ship during the training and prediction process,reducing the need for concerns about irrelevant information.We also borrowed the idea of SPP module in YOLO v4 network,and added SPP module on the basis of improved YOLO v3-CBAM3 and YOLO v4-tiny-CBAM2,in order to enhance the expressive ability of output feature map.The experimental results show that:after adding the hybrid attention mechanism,the average precision of YOLO v3,YOLO v4 and YOLO v4-tiny on the test set is improved by 0.48%,0.74%and 0.51%,respectively.After adding the hybrid attention mechanism and spatial pyramid pooling at the same time,the average accuracy of YOLO v3 and YOLO v4-tiny is improved by 0.77%and 0.62%,respectively.This shows that the hybrid attention mechanism and spatial pyramid pooling can help improve the ship detection performance of the original algorithm,and the improvement is even more helpful when used at the same time.Finally,the Faster R-CNN,YOLO x and SSD algorithms in Chapter 4 are trained on multi-class datasets and compared with the algorithms in this chapter,which not only verifies the conclusions of Chapter 4,but also proves the superiority of the improved algorithm...