Transfer classification for distinct manifestations with shared information

An object often has many distinct manifestations in computer vision, which brings a great challenge to utilizing more comprehensive information. Inspired by some biological researches about edge sensitivity and global structure priority, our key insight is to establish unified transfer classification network with shared contour information. Combining two convolutional networks with three cascaded filters, we build a unified kernel SVM classifier based on shared contour features. Two convolutional networks are used for acquiring the contour information of objects exactly. Obtained by three cascaded filters, shared edge features are used by a unified kernels SVM classifier. Our transfer classification network(TCN) is trained and tested with distinct manifestations including real photos(imagenet dataset or cifar-10 dataset) and cartoon abstracts. The model is able to extract robust contour features and achieve considerable transfer recognition accuracy(40% relative improvement to some popular convolutional models).