Research On Augmentation Technology Of Marine Ship Target Image Sample

With the improvement of computer hardware level,deep learning technology has ushered in its development period,and neural network has been developing towards the direction of more and more large-scale.In the training process of neural network,samples often play a crucial role,and the quality of data set samples directly determines the quality of training results.When we are to do the tasks such as Image segmentation or Target Recognition and so on,a large number of maritime target images are needed as data sets,but we can direct access to the marine scene pictures from the Internet is often some photographic amateur or professional staff for a particular purpose,the shooting angle and the information such as weather conditions are similar,which cannot cover all cases of the sea scene.And the number of pictures in different scenes varies greatly.There are relatively few pictures of bad weather,which direct the training is easy to result in a fitting of the neural network.The result of minority samples is not good.therefore,it needs to be augmented.In this paper,we use the following different methods for different marine scenarios to expand the scarce samples of marine scene.First,on the basis of studying the principle of style transfer and its improved algorithm,fast style transfer,this paper takes common sunny marine scene sample pictures as content pictures,dusk and cloudy pictures as style pictures,and uses Fast Style transfer to generate marine scene pictures at dusk,cloudy,foggy and rainy days.The experimental results show that the fast style transfer produces a fairly good veracity of sea scene images in dusk and cloudy sky,and can be used as a sample augmentation method for sea images in dusk and cloudy days.Secondly,a method for data augmentation of foggy weather over the sea based on atmospheric scattering model is developed.To solve the problem that it is difficult to accurately estimate a single image depth map,a data augmentation method based on image semantics segmentation and atmospheric scattering model is presented.We divide sunny pictures into three parts: sea,sky and foreground,and simulate the depth according to their structure information.Finally,the depth map is brought into the atmospheric scattering model.We generate realistic images of the foggy sea scene from the sunny sea scene picture to effectively increase the number of sample images of the foggy sea scene.Thirdly,the pix2 pix algorithm based on the Generative Adversarial Networks is used to increase the number of image samples in rainy days by training the manually made sunny-rainy pairs of image datasets.To solve the problem that the texture of the generated image differs greatly from the texture of the target in rainy days’ dataset,a method of augmenting the samples of the rainy days on sea based on the improved pix2 pix is presented.The improved algorithm replaces the CGAN structure with the GAN structure and uses the residual network as the generator to generate more realistic images of rainy sea scenes.Fourth,the study on data augmentation of marine rare images based on simulation maps is carried out.To improve the authenticity of the simulation diagram and reduce the adverse effect of the difference between the simulation diagram and the real one on the training results,a method combines with Poisson fusion and Cycle GAN is presented.This method separately uses Poisson fusion and Cycle GAN to improve the authenticity of background and foreground,and generates realistic target pictures of marine scenes from simulation images.This method can realize the purpose of training with simulation pictures,and solve the problems of ship posture,single angle and absolute scarcity in training set.