Research On Large-scale Image Classification Based On CUNet

Nowadays, along with the rapid development of the Internet, the image amount and intra-class variability both achieve explosive growth. Facing such large-scale mixed image information, how to correctly organize and manage these image data have extremely important significance. As the basis research in computer vision, image classification is the key solution to this issue.In recent years, deep convolutional neural networks have demonstrated breakthrough accuracies for large-scale image classification. Training with sufficiently large and diverse datasets, those CNN based networks successfully obtain state-of-the-art performance on visual recognition tasks. In the formulation of convolutional neural network,the convolution filter plays the central role. To learn an effective filter bank at each convolution stage, researchers are dedicated to proposing optimization algorithms. In general, previous CNN networks learn the filter bank by utilizing stochastic gradient descent method on large number of labeled images, which strictly relies on the expertise of parameter initiation and fine tuning. In addtition, such filter learning procedure will fall into local optimum, and even can not reach convergence. Besides, traditional CNN is based on supervised learning, which means that the image label is strictly required. However,nowadays, while the image scale is increasing, the image label becomes scarce, which undoubtedly brings difficulty to learn the CNN model in current application environment.Considering the aforementioned restrictions of CNN, this thesis proposes to construct a compact unsupervised network for image classification. The concrete design includes following aspects:1. Use the classical clustering algorithm K-means on the preprocessed image patches to learn the filter bank. Such filter learning is very compact because it abandons the millions of parameters initiation and fine tuning of the traditional convolutional neural network, which effectively avoids falling into local optimum. Besides, such filter learning is in unsupervised manner, which relieves the bottleneck of the scarce labeled images;2. Propose a new down-sampling method named weighted pooling. The proposed weighted pooling fully considers the different effects of all the activations in the pooling region via giving an exact weight, which contributes to improve the robustness to small image distortions;3. In the output layer, the simple histogram feature is computed, and the spatial information is fused, which improves the geometric invariance of the model. Besides,in order to remove the feature redundancy, and control the feature dimension, this thesis max pooling the adjacent blocks, which helps transfer the local histogram into global feature, and improve the robustness to image distortion.