Ship Detection Based On Light-weight Network

Target detection is an important research content of disciplines such as computer vision and artificial intelligence.It has a wide range of applications,such as license plate recognition and pedestrian detection.Among them,in the fields of maritime traffic supervision,marine environment monitoring,and suppression of illegal maritime behaviors,the detection of marine vessel targets has particularly important application background and practical significance.However,ship detection in the marine environment is susceptible to uncertain factors such as sea fog weather,complex background,ship distribution,and target distance.It is necessary to ensure the efficiency and reliability of ship detection tasks at the same time.In response to the above problems,this thesis proposes a ship target detection algorithm that takes into account detection accuracy and speed,which further improves the ship detection effect in sunny and sea fog weather environments.The specific research content of this thesis is as follows:Aiming at the ship detection algorithm,this thesis has done two researches on image preprocessing and data preprocessing.In terms of image preprocessing,in view of the problem of low ship detection accuracy in sea fog weather environment,the residual difference between foggy and non-fog images is learned through the GCANet defogging algorithm to achieve defogging processing for sea fog images.Aiming at the background interference problem in the marine environment,artificial image mask processing is used to remove the background interference of the complex environment.In terms of data preprocessing,this thesis created a new ship image data set and sea fog ship image test set,and increased the data volume and background complexity through the Mosaic data enhancement method.In addition,the Kmeans++ clustering algorithm is introduced to improve the average Io U between the network anchor boxes and the ground truth boxes.Aiming at the problem that traditional ship target detection algorithms cannot meet the actual needs of maritime tasks in real-time,this thesis proposes an improved YOLOv4 algorithm.First of all,the improved algorithm introduces the lightweight network Mobile Net V3 as the backbone network to increase the network speed.Subsequently,based on the idea of CSPNet,a new CSPbneck structure was proposed in the backbone network,which effectively improved the feature acquisition capability of the convolutional neural network and reduced the amount of network parameters and calculations.Finally,in the Neck and Head network part of YOLOv4,the SK convolution with adaptive receptive fields is introduced to further improve the detection accuracy of the algorithm while keeping the network lightweight.In order to verify the effectiveness of the improved algorithm,the commonly used performance evaluation indicators of the target detection algorithm are summarized and explained,and the lightweight network-based ship detection algorithm proposed in this thesis and the traditional target detection algorithm are compared and analyzed.The experimental results verify that the improved ship detection algorithm takes into account both the speed and accuracy characteristics of ship target detection,and improves the ship detection effect in sea fog environment.In addition,a visualized ship detection system was developed using Py Qt5,which can more intuitively display the target detection effect of ship detection,and collect the detection data in real time,and finally realize practical application tasks such as marine ship management.