it is used to demonstrate the power of convolution network. instead of classification, the convnet has the abilities to detect the object size, location, orientation,color, etc as shown by R-CNNs and YOLO. the final classification is just a combination of above factors I believe. faster R-CNN and YOLO are used to detect the object calss and also their locations. but for verification code recognization, bounding-boxes are not really need in fact, we only need to know the sequence of the chars from left to right. so I reduce the network for this specific task. this is NOT yolo, just inspired by it.