A New Method of Multi-Scale Receptive Fields Learning

Deep learning architecture has been applied in computer vision to learn features in an unsupervised manner. Thousands of features can be achieved in such manner. Furthermore, in some modified architectures, multi-scale features which contain middle layer features and output layer features, can connect to classifier. The classifier is trained using these features to predict the label of input image. The multi-scale can provide both global structures and local details, but it is prone to cause overfitting due to the expansion of features, which will make the performance degrade. In this paper, we propose a method to limit the number of features by multi-scale receptive fields (MSRF) learning. With this method, we can choose the most effective receptive fields in multiple scales. It will improve classification performance in the object recognition task. In our experiments, we compare several pre-define pooling strategies and receptive fields learning algorithm. The MSRF learning achieves the best performance among the results.