Scene Recognition Based On Deep Learning Methods

Scene recognition is one of the most important problems in the field of computer vision. Lots of practical image processing applications, such as aerial vehicles, augmented reality, and so on, are all demanding on the performance of scene recognition. Comparing with traditional image classification problem, scene recognition has a wide range of data and much more complicated input information. In this paper, two kinds of convolutional neural networks are designed for scene recognition. These networks are improved with input size and feature size. At the same time, the squared convolutional units with large area are optimized into the superposition of two different kinds of convolutional layers. This made the network deeper and improves the adaptability of the neural network. Also, the activation function in the networks is changed to PReLU function and Batch Normalization layers are added between convolutional layers. Finally,two different structures of deep convolutional neural networks are created for scene recognition.Additionally, during the process of network training, this paper presents a more agreeable method for random image cropping.Comparing to the original method, this one will maintain the input aspect ratio unchanged. Also, different kinds of image noise are added to the input layer, both performing as data augmentation. In the network testing stage, this paper uses multi-scale verification method to improve the performance of the network on the validation set.Finally, the two neural networks proposed in this paper have achieved 85.44% and 85.04% top 5 accuracy on the Places scene classification dataset, which is higher than 83.13% of WM Team, the leading team of Places 2015 challenge.