Research On Ship Target Detection Based On Deep Learning

With the booming development of deep neural networks in recent years,target detection has achieved a wide range of applications in the field of computer vision.It locates instances of target objects in images or videos,aiming to not only a continuation of classification tasks,but also the basis for advanced vision tasks such as target segmentation and target tracking.It has excellent performance in the fields of face detection,remote sensing detection,license plate detection,surveillance security,etc.Ship image is used to reflect the real scene seen by human vision by using image information acquired by visible light camera,and ship target detection is mainly for ship image.This has advanced the rapid development of intelligent ships to a certain extent.Despite the great progress in the development of target detection,there are still problems in the field of ship target detection such as small data sets and unbalanced sample classes,which lead to poor generalization ability and low detection accuracy of the model.Therefore,this thesis investigates and analyzes the ship target detection algorithm based on deep learning,and the main research contents are as follows.(1)To address the imbalance in the number of ship images of each category in the ship dataset to be detected,images of sailboats are collected and labeled in this thesis,and added to the Seaships dataset to form a new extended dataset.Based on the RetinaNet network structure,the channel attention module is added to the output layer of the residual network,and different channels are used to focus on different visual parts to strengthen the perceptual field.In this thesis,CIOU is used to take into account both the non-coincidence between the detection box and the real box and the aspect ratio of the Bounding box to improve the regression accuracy.In addition,this thesis uses Cyclical Focal Loss function to increase the number of training for fewer samples,which enhances the detection ability of the network and solves the problem of imbalance between ship target classes to a certain extent.Experiments show that on the extended Seaships dataset,the proposed method in this thesis improves the mAP index by 2.5% with the original RetinaNet.(2)To address the problem that the scale variability of target objects of ship images is too large,and also to enhance the detection accuracy of small and medium-sized targets,an improved network based on the Faster RCNN algorithm with multi-task enhancement is studied in this thesis.Firstly,ResNet50 is adopted as the feature extraction network;secondly,drawing on the feature pyramid idea,feature fusion is performed using upsampling and lateral connection,thus realizing the multi-resolution perspective analysis of images and enhancing the network’s ability to extract semantic and positional information;finally,the K-Means clustering algorithm is used to regenerate the anchor frames using the cluster idea,replacing the manually set anchor frames.Experiments on Seaships’ dataset show that the proposed method in this thesis improves the mAP metric by 4.03% over the original Faster RCNN algorithm.(3)In order to deploy the algorithm models proposed in Chapters 3 and 4 into the system,this thesis designs and implements a Python-based ship target detection system.Through functional and non-functional requirement analysis,the ship target detection system is developed based on the improved RetinaNet and multi-task enhanced Faster RCNN algorithm models,and the functional requirements of each module are realized.Users can select detection network for real-time monitoring of images and videos after login.