海大数字图像课件10.pptVIP

* * * * * * * * * * FCN对图像进行像素级的分类，从而解决了语义级别的图像分割（semantic segmentation）问题。与经典的CNN在卷积层之后使用全连接层得到固定长度的特征向量进行分类（全联接层＋softmax输出）不同，FCN可以接受任意尺寸的输入图像，采用反卷积层对最后一个卷积层的feature map进行上采样, 使它恢复到输入图像相同的尺寸，从而可以对每个像素都产生了一个预测, 同时保留了原始输入图像中的空间信息, 最后在上采样的特征图上进行逐像素分类。最后逐个像素计算softmax分类的损失, 相当于每一个像素对应一个训练样本。 * * 在基础特征网络之后又加入了全连接的分割子网，由原来的两个任务（分类+回归）变为了三个任务（分类+回归+分割）。检测和分割是并行出结果的，而不像以前是分割完了之后再做分类，结果是很 amazing 的。 主要思想：首先对图片做检测，找出图像中的ROI，对每一个ROI使用ROIAlign进行像素校正，然后对每一个ROI使用设计的FCN框架进行预测不同的实例所属分类，最终得到图像实例分割结果。? MaskRCNN优点： 1.基础网络的增强，ResNet-101+FPN 2. RoIAlign 层的加入，说白了就是对 feature map 的插值，直接的ROIPooling的那种量化操作会使得得到的mask与实际物体位置有一个微小偏移 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 10.3.3基本自适应门限(2) To obtain a good threshold, the input image can be divided into many small blocks 10.3.3基本自适应门限(3) 10.3.5最佳全局和自适应门限(1) 10.3.5最佳全局和自适应门限(2) 10.3.5最佳全局和自适应门限(3) 10.3.5最佳全局和自适应门限(4) 10.4 基于区域的分割 基本内容： 基本公式 区域生长 区域分离与合并 In sections 10.1 and 10.2 , we find boundaries between regions based on discontinuities in gray levels to partition an image into regions. In section 10.3 segmentation was accomplished via thresholds based on the distribution of pixel properties, such as gray-level values or color. In this section we discuss segmentation techniques that are based on finding the regions directly. 10.4.1 Basic Formulation 10.4.2 Region Growing Region growing is a procedure that groups pixels or subregions into larger regions based on predefined criteria. the basic approach is to start with a set of “seed” points and from these grow regions by appending to each seed those neighboring pixels that have properties similar to the seed. If the result of these computations shows clusters of values, the pixels whose properties place them near the centroid of these clusters can be used as seeds. The selection of similarity criteria depends not only on the problem under consideration, but also on the type of image data available. 10.4.2 Region Growing Another problem in