|
In recent years, convolutional neural network (CNN) has performed well in a number of image classification tasks, but it hit a bottleneck on scene recognition task, due to the multilevel semantic information in a scene. This paper is dedicated to studying the deep learning methods in scene recognition task, and making contributes to improving the classification performance in the field of scene recognition, and an effective method that captures and fuses multi-level semantic information is proposed. First of all, we compare the differences between object classification task and scene recognition task in order to apply the successful replication of CNN in object classification task to scene recognition task after resolving the differences. Then we use a multi-scale learning method to capture different scale visual features at multiple levels. In addition, on the basis of multi-scale learning, we propose a method of feature fusion at the level of category, aiming to effectively combine different scale features. The experimental results show that the success of the object classification task can be applied to the scene recognition task by our method. |
|
Keywords:Pattern Recognition and Intelligent System; convolutional neural network; scene recognition. |
|