| Computer vision technology has become an important means for intelligent ships to perceive the surrounding environment.However,in foggy and rainy weather with poor visibility,due to the scattering of suspended sea fog particles and the blocking of rain fringes,the quality of the visual images captured by the perception system,such as visibility,contrast and color,is significantly decreased,which seriously affects the high-level tasks such as visual target detection,target tracking and semantic segmentation.As a result,the accuracy is reduced and the target can not be detected or the tracking target is lost.The degradation of image quality caused by rain and fog limits the wide application of computer vision technology in the field of intelligent ships and intelligent navigation.It is necessary and important to study the fog and rain removal processing of Marine scene images.At present,there are relatively few methods to remove fog and rain for nautical scenes,and most of the existing algorithms are for land scenes.It is more difficult to remove fog and rain due to the large proportion of sea and sky,single color and few water objects in nautical scene images.In this thesis,a special data set of nautical rain and fog images is constructed,and intelligent enhancement algorithms for fog removal and rain removal of nautical images are studied based on deep learning theory.The main work completed includes:(1)Construct the nautical foggy day image data set and nautical rainy day image data set.In addition to collecting unpaired data sets of real sailing scenes,the artificial foggy day data making method is also used to form paired foggy day data sets.The artificial paired rainy day data making method of structural noise rainy day data making method and the real rendering rainy day data making method are also used to form artificial paired rainy day data sets,which can be used to better train the intelligent enhancement model of fog removal and rain removal of sailing images.(2)A navigation image defogging algorithm is proposed,which combines dark channel prior and cycle to generate adversarial network.Firstly,the dark channel prior de-fogging algorithm based on the physical model is used to decompose the image,and the transmission image and de-fogging image are output respectively.Then,the transmission image and de-fogging image are processed and judged by the cyclic generation antagonistic network,so as to generate a better de-fogging image.The comparative experimental results show that the defogging effect of this algorithm is highly consistent with the image quality evaluation index and vision,which is superior to other model methods.(3)A multistage progressive conditional countermeasure enhanced navigation image rain removal algorithm is proposed.Firstly,the multi-stage progressive rain removal module is used to remove the global rain lines from the image,and the dark channel prior defogging model is used for preliminary defogging.Finally,the conditional generation antagonistic image enhancement module is used to solve the problems of dark light,water mist and detail loss.The experimental results show that the rain removal effect of this algorithm is better than other models.(4)The weather recognition method for navigation images based on Efficient Net-B3 is designed.Methods Based on the classification of weather conditions of nautical images,the nautical images were identified as foggy or rainy images,and then input into the corresponding intelligent enhancement models of fog or rain removal to obtain clear nautical images.The intelligent enhancement algorithm for fog and rain removal of Marine images can better realize the navigation environment target recognition and tracking based on computer vision,which has important significance and application value for building a more accurate and robust ship intelligent visual perception system. |
PDF Full Text Request