For visual recognition,middle-level features provide a bridge between lower-level pixel-based information and higher-level concepts such as objects and scenes.An effective medium level representation can abstract the information directly derived from pixels at the lower level,which plays a crucial role in the subsequent classification tasks,and is robust to irrelevant information and noise.Freehand sketches is a good material for this research,and its simple and general information makes it easier for us to study presentation learning well.The main purpose of the sketch classification problem based on parts is to study the feature representation method that uses parts to represent the whole,and to try to solve the problem that the previous image classification method is not ideal in the face of intra-class change,visual Angle change,non-rigid transformation,occlusion or incompletes,etc.(1)The capsule network is an improvement on the traditional neural network structure.Its use vector instead of a scalar thoughts,and to describe the relationship between characteristics of different layers using dynamic routing algorithm,makes up for the convolution network's lack of ability to take into account the characteristics of important spatial hierarchy,and lack of rotation invariance,providing a breakthrough innovation model structure in image recognition.(2)The feature extraction method based on support vector machine provides a concise and efficient representation learning model,and with capsule network combined to explore the combination effect of traditional machine learning method and new representation learning method. |