基于深度学习的图像分割关键算法科学技术与信息专业.docxVIP

PAGE PAGE 1 基于深度学习的图像分割关键算法 摘要 随着大数据的应用和机器学的发展，人工智能逐渐从虚无走向现实。在机器人学中，机器人的学习都是基于大量数据的基础上，对大量的数据进行特征提取和特征分析，进而找到不同数据之间的相同点，进而进行判断。当处理图片时，深度学习也是通过区分图像中每隔像素的RGB或者灰度进行判断，并在像邻域内部进行特征提取。针对图像分割问题，本文从弱监督学习的角度出发进行图像分割，通过一定的规则对图像中部分物体的缺失部分进行补偿，其具体实现流程如下： 根据所需要处理的图像构建该图像的邻域图，在构建邻域时，需要参考图像和其相对应的激活映射关系。需要将像素内部及像素周边设定半径范围内像素进行连接。连接的语义相似度可以通过AN进行估算。对于单个的类而言，CAM中的稀疏激活是随着图上的随机游走将稀疏激活传递到周边语义相同的区域，边缘相似度对鼓励语义相似区域进行稀疏激活传播，同时也对边缘不相似区域惩罚稀疏激活传播。这种传播规则会一定程度的对CAM进行修改，通过修改CAM可以恢复图像的形状。这一系列过程作为一个集合，并将大量的类似集合作为训练集进行图像训练。这样一来可以获得分析大量像素处修改的CAM的最大激活相关联的类标签，进而可以进一步合成分割标签。然后通过将大量训练生成的分割标签通过SegNet网络进行二次训练，进而得出用于测试的图像分割模型。 本文的算法是在标准数据集的基础上进行了检验，根据实验结果，可以明显的验证了本文算法的合理性及有效性。 关键词：神经网络；弱监督学习；语义分割 Key Algorithm of Image Segmentation Based on Deep Learning Abstract With the application of big data and the development of machine science, artificial intelligence gradually moves from nothingness to reality. In robotics, robot learning is based on a large number of data, a large number of data for feature extraction and analysis, and then find the same point between different data, and then judge. When processing a picture, depth learning is also judged by distinguishing the RGB or grayscale of every pixel in the image, and extracting features in the image neighborhood. In view of the problem of image segmentation, this paper starts from the perspective of weak supervised learning to segment the image, and compensates the missing part of some objects in the image through certain rules. The specific implementation process is as follows: The neighborhood map of the image is constructed according to the image to be processed. When constructing the neighborhood, it is necessary to refer to the activation mapping relationship between the image and the corresponding image. It is necessary to connect the pixels within the radius set by the interior and periphery of the pixel. The semantic similarity of connections can be estimated by an. For a single class, the sparse activation in cam transfers the sparse activ