Research On The Acceleration Of Convolutional Neural Networks Via Early Decision

Currently Convolutional Neural Networks(CNN) have been widely used in the field of computer vision and it has achieved the best performance for many tasks, such as image recognition, object detection, image segmentation and so on. However, CNN is an artificial neural network with deep structure and the feed-forward propagation of CNN needs a large amount of convolution computation. Moreover, the number of the layers and the nodes in hidden layers increases as the difficulty of the task increases.Therefore, the high computational complexity limits the application of CNN in the real-time tasks, for instance, the traffic surveillance which mainly contains the detection of person and car; it also limits the application of CNN in the low-end device.This thesis presents an in-depth study of the method of accelerating CNN in the application of image recognition and object detection. The main contributions of this thesis are:1. This thesis proposes a method of accelerating CNN via early decision, aiming at the application of CNN in the binary image classification and specific object detection. This method constructs a cascade classifier based on CNN and achieving a great acceleration, taking full advantages of the discriminative features in multiple layers of CNN and the redundancy of the features. We propose a model of feature selection to select feature points across the multiple layers of CNN for each stage of the cascade classifiers. And the feature selection model balances the classification ability and computational cost of the feature points selected across the layers of CNN.We have evaluated this method on CIFAR-10, Pascal VOC2007, INRIA, TRECVID.The experiment results show that this method achieves a great acceleration with little drop in performance for the application of CNN in the binary image classification and specific object detection. Additionally, this method can be jointly used with many of the previous CNN acceleration methods to achieve a greater speedup.2. This thesis develops a system for the method of accelerating CNN via early decision. The system mainly contains: the module of extracting the training set of the feature selection model, the module of quantifying the network structure of CNN, the module of training the model of cost-sensitive feature selection and the module to construct the cascade classifiers. The method proposed by this thesis has a great time and space complexity in the training process for the CNN which has a deep and complicated structure. This system furthest reduced the time and space consumption in the training process by using some tricks.