Ship Target Detection On The Sea Surface Based On Deep Learning

China has one of the longest coastline in the world and its marine resources are abundant.For a long time,it has encountered from the challenge of neighboring countries and the strategy of returning to the Asia-Pacific of major western powers.Infrequent violations such as illegal intrusion of sea areas and confrontation with ships are frequent.The importance of national marine security has reached an unprecedented new height.For the violation of our territorial waters of the illegal ships,we need to make quick discovery,timely forensics,real-time warning.Due to the influence of uncertainties in the sea,such as illumination,fog,ship distribution,and target distance,the ship detection needs to be efficient and reliable.According to the advantages of deep learning in big data feature learning,a rapid detection method for ship targets based on deep learning was proposed.In this paper,the physical model of atmospheric scattering is introduced firstly,and the image of ship under ocean background is dehazing preprocess.The dark channel prior dehazing algorithm has the disadvantages of low processing speed and poor result on the sky region.This paper proposes to reduce the calculation of transmissivity by using the downsampling method and set the maximum threshold value of the global atmospheric light,which avoids the smear effect caused by the high atmospheric light value in some regions of the image.The image dehazing contrast experiment is designed,compare the improved dark channel dehazing algorithm with some conventional dehazing algorithms including the original algorithm;The image dehazing quality evaluation system is designed based on deep learning and color histogram.It finds out that the darkness is improved from the experiment and evaluation results.The channel dehazing algorithm has better dehazing effect.Then the YOLO network based on regressive thought is selected for real-time ship target detection.The location and recognition of ship target are combined into one net.The YOLO network is designed and improved according to the narrow and long features of the ship's target.The horizontal detection density of the network remains unchanged while the vertical detection density increased.The image is divided into 7 ×14 grids,each of them obtains 3 candidate borders including a horizontal candidate border and two vertical candidate borders.Contrast experiment is designed using improved network,original network and Faster R-CNN.The infrared image,foggy image,low-level light image and oil painting ships are detected respectively.The experimental results show that the improved YOLO network model has stronger generalization ability.Then select the Mask R-CNN network for image instance segmentation.The network uses the RoIAlign layer to solve the problem of inaccurate target positioning.For the problem of slow segmentation speed of Mask R-CNN network,an improved RoIAlign is proposed to optimize the bilinear interpolation algorithm.The improvement of the algorithm includes aligning the feature map with the center point of the target image in the pooling process,and converting the multiplication and division operations of floating-point numbers into shift operations,which reduces the amount of calculation and resource consumption.The contrast experiment was designed to compare the improved network with the original network,the MNC network and the FCIS network.The experimental results show that the improved network effect is better in segmentation speed and average accuracy.Finally,the detection system of improvd YOLO and improving Mask R-CNN network is built in windows platform.The contrast experiment is designed.The experiment environment includes occlusion targets,dense targets.The respective advantages of improved YOLO and improved Mask R-CNN networks in detection speed,classification,and positioning are verified.