In computer vision, global feature is a feature based on the extracted entire image, that is based on all pixels, a common color histogram, a shape descriptor, and the like of GIST; local feature is based on a relatively localized image blocks, i.e., based on local patches, a common feature of most of the features are local, such as SIFT, LBP and the like.
patches, a common feature of most of the features are local, such as SIFT, LBP and the like.
Under the premise of a convolutional neural network, local feature generally refers to the convolution characteristics of the network layer, (conv feature map), it retains the spatial configuration information (spatial infomation) image, features of the different positions of Feature the map, It is a partial description of different original images. And global feature refers to the spatial resolution for the top layer characterized in that the network-wide connection layer 1 * 1, this time could not find the corresponding local image, corresponding to the whole FIG. Considering the connection itself is full convolution 1 * 1, can be simply understood, feature map when the resolution is greater than 1 * 1 when the conv or fc feature is local feature, after either down to 1 * 1 fc or conv should all be considered global feature. In an example AlexNet, conv1-5 a local feature, fc6-8 Laid is global; NIN or if network inception, before doing global average pooling (GAP) is a local feature, then finish GAP becomes global feature .Knowing almost turn: https: //www.zhihu.com/question/31039746/answer/50516028