Research On Image Fusion And Segmentation Method Based On Deep Learning

Image fusion and image segmentation are important research contents in the field of computer vision.Traditional image fusion methods require manual setting of features and fusion criteria to complete the fusion task,however,traditional image segmentation methods need to use structural prior information,which does not consider the influence of potential feature distribution sufficiently and separates two crucial algorithm design procedures:feature extraction and classifier designing,limiting their performance to some extent.Therefore,this thesis solves this problem by using deep learning method,and designing deep convolutional neural network method to study multi-focus image fusion and colon gland segmentation method.The main research contents are as follows:1.we propose a deep learning method based on the spatial pyramid pooling(SPP).First,we design a Siamese network and replace the average pooling with SPP to learn the features of multi-focus images.Then,to train the network effectively,we synthesize a large-scale multi-focus image dataset with ground truth through a Gaussian filter.Given a pair of multi-focus image as input,the trained model can generate a score map indicating the focus property of source images.Moreover,to further enhance the fusion effects,we segment the score map into a binary mask image,which is refined using morphological technique.Finally,the fused image is gained by employing dot multiplication operation between source images and the refined binary mask image.Experimental results reveal that the average quantitative score on test images achieved by the proposed method is increased by 0.005%.2.Biomedical image object segmentation plays a significant role in both the intelligent medical diagnosis and clinical practice assessment.Previous methods utilize the hand-crafted features and prior structural information to fulfill segmentation task,which does not consider the influence of potential feature distribution sufficiently and separates two crucial algorithm design procedures:feature extraction and classifier designing,limiting their performance to some extent.To overcome these issues,in this thesis,we propose a method based on deep dense convolutional neural network and focal loss.First,to harness the ground truth information of colon gland adequately,we formulate the instance segmentation as a multi-task learning problem,learning the features of gland object and contour simultaneously,after which we leverage deep dense convolutional neural network,which integrates all the low-level structural and high-level semantic information,to extract the features in inputs efficiently.To migrate the class imbalance existing in training dataset,we replace common cross entropy objective with focal loss.Once trained,given colon gland images as inputs,the well-trained model output two high-quality confidence maps to indicate the score for each pixel belongs to objects or contours.To further enhance the instance segmentation effect,we harness the morphology and convolutional conditional random fields techniques to refine the confidence maps to obtain the final segmentation results.Results demonstrate that our method outperforms other popular methods in terms of both visual perception and three quantitative metrics,thereby revealing superior performance.