| Image segmentation is one of the main steps of image processing,which is mainly based on gray-scale and color-related features to divide the image to form a certain continuous area and extract the target object.Due to the uncertainty of the image,especially the problem of “same object with different spectrum,same spectrum with foreign object” and noise interference in remote sensing images,traditional segmentation methods often fail to obtain the ideal segmentation effect,which has a negative impact on the subsequent target recognition.In order to solve the problem of ambiguity and uncertainty of image pixel attribution,some scholars have introduced the method of fuzzy mathematics.Fuzzy c-means(FCM)clustering method has obvious advantages in the field of image segmentation,such as fast convergence speed,simple operation,and has received widespread attention,but it is still sensitive to noise,and the segmentation accuracy still needs to be further improved.As one of the representative deep learning algorithms,convolutional neural networks can perform large-scale task-driven feature learning from big data.However,typical deep learning is a completely deterministic model,which is greatly affected by data uncertainty.Therefore,stable,robust and accurate image segmentation still faces greater challenges.At present,there are image processing algorithms that combine fuzzy theory and deep learning,but they are rarely used in the field of image segmentation,and there are still many problems in these algorithms that need to be studied and solved.Therefore,this paper mainly studies the description of the uncertainty of image data in the image segmentation process,the improvement of fuzzy clustering algorithms,and the combination of convolutional neural networks and fuzzy theory in deep learning.The main research work is summarized as follows:1.Since the contextual information between pixels is of great significance to the anti-noise and accuracy of the image segmentation algorithm,the reliability of the spatial context is introduced,and a new fuzzy C-means clustering algorithm is proposed for image segmentation: Effectively model the spatial context to improve the anti-noise performance of the clustering algorithm,and study a new reliability fuzzy metric,so that the clustering algorithm can better balance detail preservation and denoising,so as to obtain more accurate segmentation result.The experiment selects three types of data:artificial synthetic image,traffic sign image and remote sensing image to test the performance of the clustering algorithm.The results show that the proposed algorithm can effectively suppress intra-class heterogeneity and inter-class isomorphism in the image segmentation process,can improve the image pixel separability,and effectively preserve the edge details of the image.2.The concept of fuzzy learning is introduced into deep learning to overcome the shortcomings of fixed representation and improve segmentation accuracy.Image data usually has a large uncertainty,which is related to many factors such as imaging equipment,which makes the traditional segmentation method greatly restricted.The existing semantic segmentation methods represented by deep learning have made breakthrough progress.The deep convolutional neural network is a completely deterministic model and cannot describe the uncertainty of remote sensing images well.Therefore,a new deep neural network combined with fuzzy logic unit is proposed,which integrates convolution unit and fuzzy logic unit.The convolution unit is used to extract discriminative features with different proportions,so as to provide comprehensive information for pixel-level image segmentation.The fuzzy logic unit is used to deal with various uncertainties and provide more reliable segmentation results,and each unit processes the feature map at a specific image scale through the Gaussian blur function.Through hyperspectral data sets tested,the proposed algorithm has higher segmentation accuracy and better performance. |
PDF Full Text Request