Abstract:

There are two primary problems  in scene labeling: how to produce good internal representations of the visual information and how to use contextual information efficiently. To solve the two problems, we present a scene labeling algorithm for multiscale deep networks based on deep learning, a supervising model. Unlike traditional multiscale methods, the model contains two deep convolutional networks: one takes the global information into account and extracts the lowlevel features of the largescale image; and the other one combines the local information of the image with the low level features, and obtains a set of dense and complete image features, thus a powerful representation that captures texture, color and contextual information is achieved. Compared with many standard approaches, the proposal does not depend on segmentation technique or any task specific features. Our approach yields good performance on the Stanford Background Dataset.

Key words: scene labeling;multiscale deep network;supervised learning;deep convolutional network