Research On Fouling Image Segmentation Based On Level Set And Convolutional Neural Network

With the development of science and technology,image acquisition equipment has gradually become popular,and the cost of acquiring images has become lower and lower.Therefore,in the field of computer vision,the requirements for image segmentation have become higher and higher.Especially when facing fouling images,the segmentation model is required to accurately segment the contours of the target.Fouling images refer to the loss of pixel information of the target in the image or the image contains various noise pollution,etc.The cause of the loss of pixel information may be the target is blocked by other objects or some information is lost during the image transmission.The traditional level set-based segmentation method has been accepted by scholars through the curve evolution theory and extended research.It can get good segmentation results when the image target is simple and has no interference information such as defacement,because this type of method can extract the target at the pixel level.However,when facing the fouling image,the pixel information is lost,which makes it impossible to extract effective features at the pixel level,and cannot obtain accurate segmentation results,and this type of segmentation model often requires a wealth of artificial prior knowledge to design methods of extract feature related,so it has many limitations in image segmentation.In recent years,with the introduction of convolutional neural network technology into computer vision,good results have been achieved.In particular,neural networks such as U-Net have achieved significantly better results than previous models in image segmentation in certain fields.A large number of samples are trained to fit the mapping relationship between the input image and the ground truth,so as to obtain a good segmentation result,but the higher accuracy of this type of method is based on a large amount of training data.When the sample data is small,the correlation mapping relationship cannot be accurately fitted,resulting in the final failure to obtain a good segmentation result.Based on the merits of the above two methods,this paper combines the two into a framework,and uses their respective advantages to achieve the purpose of accurate segmentation of the fouling image.First,in the face of small sample fouling images where the target is occluded or part of the pixel information is missing,it is proposed to add the constraint information of the target contour in the CV model based on the level set method to the U-Net network to assist its feature extraction process.It not only considers the mapping of the corresponding pixels between the input image and the ground truth,but also pays attention to the similar information between the same type of targets,and finally can complete the effective segmentation of the fouling image under the premise of small samples.Secondly,in the face of noise-contaminated fouling images,a segmentation method that automatically separates the noise field is proposed.The current noise image segmentation methods assume that the noise distribution is known or observable,it does not model the noise field contained in the noise image.Based on this problem,a method combining convolutional autoencoder and traditional offset field model is proposed to segment noisy images.Take a noisy image as an input image,superimpose a layer of random noise on it,and obtain a convolutional autoencoder that can remove superimposed noise through training,and use it to remove the noise field in the original noisy image.At the same time,the denoising convolution kernel trained in the convolutional autoencoder is used as the filter in the offset field model to achieve the purpose of segmentation of the noisy image.While accurately removing the noise,it also avoids the error caused by the artificial design of the denoising filter in the traditional offset field model.Finally,by designing a series of comparative experiments to verify the effectiveness of the model proposed in this paper for segmentation of fouling images,this article mainly combines the traditional segmentation method with the deep learning framework,utilizing the respective advantages of the two,a higher accuracy rate of segmentation of defiled images is achieved.