Deep Learning Literature Reading notes (1)

Blink of an eye has been two, suddenly want to read the previous literature summary and summed up with you to share, stay as a souvenir, convenient for later reference.

1. Deep tracking: Visual tracking of differential feature learning via convolutional networks (deeptrack:learning discriminative Feature Representations by convolutional neural Networks for visual Tracking) (English, conference papers, 2014, ei Search)

An article using convolutional neural network for target tracking can not only use CNN as a pattern recognition, but also can do target tracking, after all, it is essentially a means of feature extraction.

2. Research on vehicle identification method based on deep learning (Chinese, periodical, 2015, net)

It is a new problem to use the traditional CNN for vehicle identification, advanced traffic standard location extraction, train in CNN, and finally use support vector machine to classify. Experimental hardware configuration: Clock 2.80ghzcpu,2g memory, not used to GPU acceleration.

3. Method of defect recognition of radiographic images based on deep Learning Network (Chinese, periodicals, 2014, knowledge net)

Using CNN directly for radiographic defect detection, old methods and new problems, the CNN structure is described clearly, suitable for CNN.

4. Deep learning and its new progress in target and behavior recognition (Chinese, journal, 2014, net)

This paper mainly summarizes the structure and application progress of self-coder and restricted Boltzmann machine in deep learning, summarizes the comprehensive and authoritative, and describes the principle and improvement progress of both, and points out that "deep learning is a multilayer depth structure, the signal is propagated in this multilayer structure, and finally the signal expression is obtained. Learning to multi-layered nonlinear functional relationships, better modeling of visual information, is worth reference.

5. Significant target detection based on hyper-pixel convolutional neural network (Super Cnn:a Superpixel wise convolutional neural Network for salient object detection) (English, Journal, 2015 , IEEE Search)

CNN in the field of target detection, the first image of the super-pixel segmentation, get three sequences (hyper-pixel sequence, a space kernel matrix, a range of core matrix), and then three sequences into three convolutional networks for training, to achieve the multi-channel input of CNN, the significance of the main by the super-pixel to reflect. However, the training speed of the article experiment is super slow, it takes 4 days to 6 days.

6. Cascading features learn face Recognition (Joint Feature Learning for recognition) (English, Journal, 2015, IEEE Search)

The basic idea is "image block + deep learning", and the implementation of unsupervised training, similar to the sparsity algorithm, is an abstract reference to deep learning, that is, only reference the concept of "depth", the article conveys an important message: the so-called Deep network, is in a layer of search mapping matrix, Each coefficient in the matrix is a parameter to be learned.

7. Robust pedestrian count using sparse representations and random mappings (robust people counting using sparse representation and random projection) (English, Journal, 2015, IEEE Search)

The use of CNN for pedestrian detection, to achieve the semi-supervised training, in fact, is the sparsity algorithm, no block thinking, no super-pixel thinking.

8. Combined deep Learning pedestrian detection (Joint Deepin Learning for Pedestrain Detection) (English, conference papers, 2013, IEEE Search)

Use CNN for pedestrian detection. In fact, the use of CNN in pedestrian detection is relatively large, and the improvement of a variety of ways. Based on the deformation model in pedestrian detection, this paper analyzes the parts of human body and points them out, and it is an abstract improvement to CNN by using the theory of deep learning model in the deformation layer.

9. Deep-learning face representation based on multi-category prediction (Deepin learning faces representation from predicting 10000 classes) (English, Conference, 2014, IEEE Search)

The Chinese University of Hong Kong Deepid generation algorithm article, the use of CNN for face authentication, first face detection of the image of the key points, and the block, the CNN structure using the Relu activation function, and sampled after two layers of mapping results for feature fusion, combined with the combined Bayesian classifier, Experiments were conducted on celebface and LFW databases to achieve 97% accuracy.

10. Deep learning face representation based on the recognition of joint human face authentication (Deepin learning faces representation by Joint identification verification) (English, Conference, 2014, NIPS)

The Chinese University of Hong Kong deepid second-generation algorithm article, based on the generation algorithm to add the verification signal to improve, and by confirming the signal and the verification of the combination of signals to improve inter-class dispersion, reduce intra-class dispersion, and a variety of confirmation of the signal is compared, in the experimental database has been expanded, the correct rate of 99.15%

11. Deep Learning Facial expression sparsity, selectivity and robustness (learning face representation is sparse,selective and robust) (English, Conference, 2014, CVPR)

The Chinese University of Hong Kong deepid three-generation algorithm article, the output layer from 160 to 512 dimensions, the expansion of the experimental database (290,000 training samples), each layer of the neural network to add a verification signal (two-generation algorithm only validated the second layer), while the first two generations of the algorithm to summarize, with the experiment to demonstrate the sparsity of CNN output, Selectivity and robustness, explaining the meaning of CNN data.

Copyright NOTICE: This article for Bo Master original article, without Bo Master permission not reproduced.

Deep Learning Literature Reading notes (1)