Deep Learning Based Weakly Supervised Classification Method And Application

With the increasing demand of intelligent data analysis and processing in the big data era, training classifiers in a supervised way with large scale manual labeled data is out of time. Weakly supervised classifier, i.e. semi-supervised and unsupervised, becoming one of the key method of overcoming this problem. However, most traditional weakly supervised methods are only suitable for small scale data and without the function of mapping non-linear feature space. Therefore, deep learning technique provide a new solution for these problems with its natural advantages. Based deep learning technique, this paper focus on weakly supervised classification method and apply them into image segmentation and gait recognition successfully. The main contributes and novelty are concoluded as followed:1. Proposed a unsupervised clustering algorithm based on deep auto-encoders. With basic auto-encoder structure, automatic clustering are achieved in feature space. With internal and external clustering restrains, better clustering result are reached. After iterations, the cluster distribute becomes more conpact. Experiments on 4 datasets demonstrating proposed method is effective.2. Proposed a semi-supervised clustering algorithm based on convolutional neural networks, containing 3 convolutional layers,3 fully-connected layers for classification and 1 fully-connected layer for reconstruction, and corresponding two restrains, namely the image reconstruction restrains and image classification restrains. This method is varified on CIFAR-10 which is one of the most challenging database for clustering, experimental results with two restrains shown proposed method is valid.3. Proposed a human image segmentation method based on convolutional neural network in an image by image manner. For the lack of labelled samples, several data augmentation techniques and drop-out are adopted to avoid over-fitting. Comapred with pixel by pixel manner, proposed method speed up 10,000 times.4. Proposed a gait recognition method merging segmentation and recognition into one framework. It takes an automatic label generating technique, jointly learning the segmentation and recognition parts, recognizing gait in real scene, experimental results showing this method is much better than traditional methods.