An Object Recognition Method Based on the Improved Convolutional Neural Network

摘要

As is well known that object recognition is a key problem in computer vision, and in this paper, we aim to propose a novel object recognition method using an improved convolutional neural network. Convolutional neural network (CNN) has the ability to learn rich features from the training data and a convolutional network is composed of convolution and pooling. Particularly, the convolutional layers is used to extract features, such as orientated edges and corners, and the averaging and sub-sampling layer is utilized to reduce the precision of the feature map. As the traditional convolutional neural network can only extract features of the same scale in each convolutional layer, hence, it is not suitable to be exploited in variable-scale objects recognition. Therefore, in this paper, we propose an improved convolutional neural network to extract multi-scale features at the highest convolutional layer and provide a method to learn the weights in the proposed CNN as well. To test the effectiveness of the proposed algorithm, the CIFAR-10 dataset, and the Neovision2 Tower Dataset are utilized in the experiment. Experimental results demonstrate that, compared with CNN-HLSGD and SCNN, our proposed approach can significantly improve the accuracy of object recognition than other models.