A new privacy attack network for remote sensing images classification with small training samples.

Solving overfitting problems of privacy attacks on small-sample remote sensing data is still a big challenge in practical application. We propose a new privacy attack network, called joint residual network (JRN), for deep learning based privacy objects classification of small-sample remote sensing images in this paper. Unlike the original residual network structure, which add the bottom feature map to top feature map, JRN fuses the bottom feature map with top feature map by matrix joint. It can reduce the possibility that convolution layers extract the noise of training set or consider the inherent attributes of training set as the whole sample attributes. A series benchmark experiments based on GoogleNet model have been enforced and finally, we compare the model process output and the classification accuracy on small-sample data sets. On the UCMLU data set, the GoogleNet-Feat model which is integrated with JRN is 1.66% higher of accuracy than the original GoogleNet model and 1.87% higher than the GoogleNet-R model; on the WHU-RS dataset, GoogleNet-Feat model is 1.04% higher than the GoogleNet model, and is 3.12% higher than the GoogleNet-R model. Compared with the contrast experiments, the classification accuracy of GoogleNet-Feat is the highest when facing the overfitting problems resulting from the small samples.

A new deep learning model for assisted diagnosis on electrocardiogram.

In order to enhance the accuracy of computer aided electrocardiogram analysis, we propose a deep learning model called CBRNN to assist diagnosis on electrocardiogram for clinical medical service. It combines two sub networks which are convolutional neural network (CNN) and bi-directional recurrent neural network (BRNN). In the model, CNN with one-dimension convolution is employed to extract features for each lead of ECG, and BRNN is used to fuse features of different leads to represent deeper features. In the training step, we use more than 40 thousand training data and more than 19 thousand validation data to obtain the optimal parameters of the model. Besides, by validating our model on more than CCDD 120,000 real data, it achieves an 87.69% accuracy rate, higher than popular deep learning models such as CNN and ResNet. Our model has better accuracy than state-of-the-art models and it is also slightly higher than the average accuracy of human judgement. It can be served for the first round screening of ECG examination clinical diagnosis.

Multi-Path Dilated Residual Network for Nuclei Segmentation and Detection.

As a typical biomedical detection task, nuclei detection has been widely used in human health management, disease diagnosis and other fields. However, the task of cell detection in microscopic images is still challenging because the nuclei are commonly small and dense with many overlapping nuclei in the images. In order to detect nuclei, the most important key step is to segment the cell targets accurately. Based on Mask RCNN model, we designed a multi-path dilated residual network, and realized a network structure to segment and detect dense small objects, and effectively solved the problem of information loss of small objects in deep neural network. The experimental results on two typical nuclear segmentation data sets show that our model has better recognition and segmentation capability for dense small targets.

A sparse deep learning model for privacy attack on remote sensing images.

Deep learning tools have been a new way for privacy attacks on remote sensing images. However, since labeled data of privacy objects in remote sensing images are less, the samples for training are commonly small. Besides, traditional deep neural networks have a huge amount of parameters which leads to over complexity of models and have a great heavy of computation. They are not suitable for small sample image classification task. A sparse method for deep neural network is proposed to reduce the complexity of deep learning model with small samples. A singular value decomposition algorithm is employed to reduce the dimensions of the output feature map of the upper convolution layers, which can alleviate the input burden of the current convolution layer, and decrease a large number of parameters of the deep neural networks, and then restrain the number of redundant or similar feature maps generated by the over-complete schemes in deep learning. Experiments with two remote sensing image data sets UCMLU and WHURS show that the image classification accuracy with our sparse model is better than the plain model,which is improving the accuracy by 3%,besides, its convergence speed is faster.

A new graph drawing scheme for social network.

With the development of social networks, people have started to use social network tools to record their life and work more and more frequently. How to analyze social networks to explore potential characteristics and trend of social events has been a hot research topic. In order to analyze it effectively, a kind of techniques called information visualization is employed to extract the potential information from the large scale of social network data and present the information briefly as visualized graphs. In the process of information visualization, graph drawing is a crucial part. In this paper, we study the graph layout algorithms and propose a new graph drawing scheme combining multilevel and single-level drawing approaches, including the graph division method based on communities and refining approach based on partitioning strategy. Besides, we compare the effectiveness of our scheme and FM(3) in experiments. The experiment results show that our scheme can achieve a clearer diagram and effectively extract the community structure of the social network to be applied to drawing schemes.