LAUSR

Abstract With the growth of deep learning, object recognition has received increasing interests and its accuracy has been improved significantly in the past few years, However, high-quality recognition largely depends on a large number of learning instances. If the number of learning instances is reduced, it’s difficult to maintain realistic recognition accuracy. Moreover, traditional methods usually don’t consider the semantic relationship between different regions. Actually, semantic constraint would contribute to improve the recognition accuracy effectively. Aiming at the problems above, we proposed one semantic constraint based object recognition method. On the one hand, instance-based transfer learning model could make use of learning instances of other categories to maintain realistic recognition accuracy. On the other hand, semantic constraint between different regions simulated as joint entropy is used to recognize target object more accurately. At last, adequate experiments using a large number of images show that our model not only could reduce the number of learning instances but also could achieve realistic recognition.