Research On Object Recognition In Natural Scene

In recent years, with the rapid development of the network, the amount of images on the Internet has been growing swiftly. How to make the computer own the ability to recognize the salient object in an image becomes an imperative demand. At present, popular object recognition algorithms can be mainly divided into two types: one by manually extracting features, and the other through deep learning method. In this paper, we respectively study the above two algorithms and put forward SIFT and SURF fusion features object recognition algorithm and DCNN recognition algorithm. Furthermore, as DCNN can't learn some good artificial features, we propose a novel DCNN model combined with Gabor filters to strengthen the learing of texture information. The main work and innovations are as follows:(1) Propose an object recognition algorithm based on SIFT and SURF itegration features. Traditional artificial feature object recognition method firstly extracts the SIFT features of images, and then uses the standard bag-of-words model to describe features. Finally, it utilizes a SVM classifier for image classification. Although the SIFT algorithm shows robustness on scale change and rotation change, its stability is still worse than SURF in terms of image blurring and brighteness change. Besides, standard bag-of-words model neglects SIFT feature positional information. Based on the above facts, we propose a SIFT and SURF itegration features object recognition algorithm. Through the constract experiments of different data sets, we prove that our presented algorithm is superior to the traditional one.(2) Propose an object recognition algorithm based on DCNN. The traditional feature extraction algorithm can only fetch low-level features, which can't form intrinsical description of the data set. Therefore, we study the deep learning algorithm and design a novel deep convolutional neural networks named 4CS-DCNN. It achieves 80.27% recognition rate on the ImageNet10 data set, significantly outperforming the SIFT and SURF itegration features object recognition algorithm.(3) Propose a DCNN object recognition algorithm combined with Gabor feature. The DCNN can't learn some good artificial features, such as texture feature. Therefore, we add Gabor feature to the DCNN so as to enhance the learning of texture feature. And ultimately it achieves 81.53% recognition rate on the ImageNet10 data set, yielding 1.30% promotion compared with the previous 4CS-DCNN model.