Online memory learning for active object recognition

Traditional CNN-based recognition algorithms are trained for limited labeled data, which may not perform well in a different environment due to the lack of adaptivity of the CNN networks. So the traditional CNN-based recognition algorithms can not play a good role in robot applications because the robots have to work in different environments. However, the robot can continuously perceive new images during its mission. These images contain lots of environment-related features but lack of labels. So the robots must learn the environment-related features adaptively with unlabeled data to further improve the performance of CNN-based recognition algorithms. We call this ability as active object recognition (OBR). In this paper, we designed a dynamic memory structure (DMS) which can adaptively learn the environment-related features online and embedded DMS into a VGG-16 network to implement active object recognition. We also evaluate our dynamic memory network of CIFAR-10 and CIFAR-100 classification dataset. The results show that by learning environment-related features, dynamic memory network achieves a better performance on classification accuracy. More importantly, the network can have the ability to improve itself while many times testing.