文献类型：期刊文献

中文题名：破损图像子区域自适应划分标注方法仿真

英文题名：Simulation of adaptive segmentation method for broken image sub-region

作者：杨柳[1]

第一作者：杨柳

机构：[1]河南财经政法大学文化传播学院

第一机构：河南财经政法大学文化传播学院

年份：2019

卷号：0

期号：4

起止页码：461-464

中文期刊名：计算机仿真

外文期刊名：Computer Simulation

收录：CSTPCD;;北大核心:【北大核心2017】；

语种：中文

中文关键词：破损图像;子区域;划分标注

外文关键词：Damaged image;Subregion;Partition annotation

摘要：针对当前图像标注方法中存在准确率与召回率低的问题,提出基于多示例学习的破损图像子区域自适应划分标注方法。利用模糊增强算法计算图像模糊隶属度,对得到的图像隶属度进行非线性变换增强图像对比度,通过隶属度的逆变换获取整体得到增强的破损图像。在对数域上分离整体增强后图像中大特征和小细节,并采用二次函数分解图像中的大特征,得到增强细节后的破损图像。将增强后图像代入图像分割中,利用相邻像素之间差分计算得到图像中具有梯度突变的点,利用形态学处理突变点,并确定初始分割阈值。将互信息量当作目标函数,通过计算分割图像和原始图像互信息量,得到最佳分割阈值,实现破损图像分割。经图像分割,利用多示例学习法与多样性密度法相结合的方式将训练图像集合中图像划分成正包、负包,利用计算图像各分割块有关各标注自身多样性密度的得分值实现图像子区域自适应划分标注。实验结果表明,该方法标注准确率与召回率均较高。
Because the current image annotation method has low correct rate and recall rate,this article puts forward a method of adaptive partition annotation for subregion in damaged image based on multi-instance learning method.The fuzzy enhancement algorithm was used to calculate the fuzzy membership degree of image,and then obtained image membership degree was nonlinearly transformed to enhance the image contrast.After that,the inverse transform of membership degree was used to obtain the overall enhanced damaged image.The big features and small details in the overall enhanced image were separated on the logarithmic domain.Meanwhile,the quadratic function was used to decompose the big features in image,so as to obtain the damaged image with enhanced detail.Then,the enhanced image was introduced into the image segmentation,and the difference calculation between adjacent pixels was used to get the point with gradient mutation in image.The mutation point was processed by morphology and the initial segmentation threshold was determined.In addition,the mutual information was used as the objective function.By calculating the mutual information between the segmented image and the original image,the optimal segmentation threshold was obtained and the damaged image segmentation was achieved.Through image segmentation,the multi-instance learning method and the diverse density method were used to divide the image in the training image set into positive packet and negative packet.Finally,the score of diversity density of each image segment was calculated to achieve the adaptive partition annotation of subregion in image.Experimental results show that the method has higher tagging accuracy and recall rate.