Mastering Computer Vision with TensorFlow 2.x

Feature extraction

Feature extraction is, grouping of similar features, such as edges, corners, and lines, into feature vectors. Feature vectors reduce the dimensionality of the image from 227 x 227 (~51,529) to 4,096, for example. Each region proposal, irrespective of its size, is first converted into a size of 227 x 227 by dilation and warping. This is required as the input image size for AlexNet is 227 x 227. 4,096 feature vectors are extracted from each region using AlexNet. The feature matrix is 4,096 x 2,000, as we have 2,000 region proposals for each image.

In principle, R-CNN can take any CNN model (such as AlexNet, ResNet, Inception, or VGG) as input as long as the input image size is modified to fit the network's image size. The ...