Research On Ship Detection Algorithm In UAV Aerial Image

With the development of mini-UAVs equipment and image processing algorithms’ progress,the time is ripe to use aerial images to detect ship targets in the inland waterway and then help maritime departments improve water traffic supervision’s ability.However,it is difficult to detect the object in UAV aerial images because of the small size of the ship target,the large scale change,and the difficulty of locating the object in a complex background.Given the above three difficulties,this paper uses UAV aerial ship image data to study the one-stage object detection algorithm based on deep learning.The new algorithm is constructed by improving the feature extraction network,detection frame regression strategy,and bounding box regression loss.These methods improve ship detection performance.First of all,introducing the construction of a convolution neural network and ship detection algorithm.This paper introduces those two modules’ structure,aiming to understand each module’s internal structure and lay a foundation for designing a new algorithm.The convolutional neural network comprises the input layer,hidden layer,and output layer,which realizes input data processing,feature extraction,and output classification results.The ship detection algorithm is an extension based on the former,which consists of a feature extraction network,multi-scale feature fusion module,detector,and loss function.Secondly,the influence of anchor frame positioning on ship object detection is analyzed.This paper studies three types of one-stage ship detection algorithms,SSD,YOLOv3,and Retina Net,based on anchor frames.After analyzing the network structure,the location of the detection frame,and the loss function of these algorithms,through the experimental test,the defects of using Focal Loss to classify the advantages and the regression of anchor frame are determined.Then the research on the detection module without anchor frame is introduced.Then,the detection algorithm based on anchor-free regression strategy,Corner Net,and feature selective anchor-free module(FSAF)are studied.The former constructs the detection box selection module based on corner regression and the latter uses the center point to determine the location of the detection box.The results show that FSAF is more concise and mature in positioning mode,and it has better performance in detection accuracy and detection speed.Finally,based on the above research,this paper proposes a one-stage ship detection algorithm,Fovea SDet,using the anchor-free frame regression strategy.First,the feature extraction network’s input module and block structure improve small object detection accuracy.Second,Fovea Head constructs the anchor-free frame regression strategy under multi-scale features to improve the influence of scale change on ship detection.Last,complete IOU(CIOU)is introduced,improves frame positioning precision,and keeps the better detection frame positioning ability in complex background.Compared with other algorithms,Fovea SDet has obtained the highest detection accuracy,proving its superiority and the effectiveness of each module improvement.