Search

In recent years, convolutional neural networks (CNNs) have been successfully applied in the field of image processing, and have been deployed to a variety of artificial intelligence systems. However, such neural models are still considered to be “black box” for most tasks. Two of fundamental issues underlying this problem are as follows: 1. What type of knowledge learned by a neural network in a task was not understandable and predictable? 2. The decision made by a neural model was generally not evaluable. Like neural coding in the brain, some neurons only participated in encoding a particular task. Inspired by this, in this paper, we propose a method to modify traditional CNN models into understandable CNNs, to clarify the information coding in high conv-layers of CNNs and further evaluate the decisions made by a neural model. In our understandable CNN models, each neural node (feature map) in a selected conv-layer was assigned to participate in encoding only one class in a classification task. Our models use the same training data as ordinary models without the need for additional annotations for supervision. We applied our method to the ResNet and DenseNet models. The experiments showed that new models can learn the information coding mode that we expected in an image-recognition task, and, using the pre-assigned coding mode, we can interpret why a neural model makes a right or wrong decision, which decisions are credible, and which are not.

It is well known that the number of users is increasing rapidly in online social networks. People are linked through various types of social relationships. Detecting the type of social relationships is fundamental to improve performance on many applications in social networks. Existing studies mainly focus on predicting social relationships in the same network based on its own abundant information. However, there are few works on predicting social relationships across different networks. In this paper, we study a tri-training model to predict the labels of social relationships across different networks based on BP neural networks. Firstly, we aim to obtain the hidden common characteristics between two different networks. Depending on these hidden common characteristics, unlabeled target samples are assigned pseudo labels by training two BP neural networks. Then we capture the special features for the target network by analyzing the structural characters. Finally, combining with the special and common features, we optimize the third BP neural network by training the pseudo labeled target samples. After training, we can use the third optimized BP neural network to predict the types of social relationships in the target network. We evaluate the model and compare it with some existing methods on six real online social networks. The experimental results indicate that our model can effectively exploit the unlabeled target sample and outperform the existing methods on multiple metrics, and also show that the special features in the target network can help to enhance the learning performance.

Deep learning technologies have played more and more important roles in Computer Aided Diagnosis (CAD) in medicine. In this paper, we tackled the problem of automatic prediction of Alzheimer's Disease (AD) based on Magnetic Resonance Imaging (MRI) images, and propose a fully unsupervised deep learning technology for AD diagnosis. We first implement the unsupervised Convolutional Neural Networks (CNNs) for feature extraction, and then utilize the unsupervised predictor to achieve the final diagnosis. In the proposed method, two kinds of data forms, one slice and three orthogonal panels (TOP) of MRI image, are employed as the input data respectively. Experimental results run on all the 1075 subjects in database of the Alzheimer's Disease Neuroimaging Initiative (ADNI 1 1.5T) show that the proposed method with one slice data yields the promising prediction results for AD vs. MCI (accuracy 95.52%) and MCI vs. NC (accuracy 90.63%), and the proposed methods with TOP data yields the best overall prediction results for AD vs. MCI (accuracy 97.01%) and MCI vs. NC (accuracy 92.6%).

Having a multitude of unlabeled data and few labeled ones is a common problem in many practical applications. A successful methodology to tackle this problem is self-training semi-supervised classification. In this paper, we introduce a method to discover the structure of data space based on find of density peaks. Then, a framework for self-training semi-supervised classification, in which the structure of data space is integrated into the self-training iterative process to help train a better classifier, is proposed. A series of experiments on both artificial and real datasets are run to evaluate the performance of our proposed framework. Experimental results clearly demonstrate that our proposed framework has better performance than some previous works in general on both artificial and real datasets, especially when the distribution of data is non-spherical. Besides, we also find that the support vector machine is particularly suitable for our proposed framework to play the role of base classifier.

Histogram based image enhancement techniques are widely used for performing contrast enhancement in images. However, most histogram based image enhancement methods have insufficient capability to freely tune the brightness and contrast of enhanced image. In this paper, two novel histogram based image enhancement algorithms are proposed. The proposed algorithms provide the way to control the brightness and contrast of enhanced image by adjusting two parameters. The principles for parameter selection are also discussed in this paper. Experimental results demonstrate a better performance of the proposed methods in both perceptual quality and image quality assessment metrics than the existing histogram based methods.

Transform coding plays a very important role in image and video compression. Discrete Cosine Transform (DCT) is used as standard scheme (i.e. JPEG) in lossy image compression. Consequently, integer Discrete Cosine Transform (iDCT) is presented to achieve lossless compression for the compatibility of JPEG. Presently, with the investigation of new and well performed image transform techniques, there is an undeniable need for novel transform coding technologies to improve the compression rates and reduce computational complexity in the field of transform based lossless image compression. Discrete Tchebichef Transform (DTT) is a potentially unexploited orthogonal transform, and has shown a number of valuable properties like energy compaction and recursive computation. It has been preliminarily introduced in lossy image compression and shown the superiority in the compression rates. However, the DTT has not been investigated in lossless image compression. In this paper, we study DTT and matrix factorization theory firstly, and then factorize the N×N DTT matrix into N+1 single-row elementary reversible matrices (SERMs) with minimum rounding errors. On this base, we introduce a novel algorithm, named integer DTT (iDTT), to achieve integer to integer mapping for efficient lossless image compression. A series of experiments are carried out and results show that the proposed iDTT algorithm not only has higher compression ratio than iDCT method, but also is compatible with the widely used JPEG standard. A framework of lossless image compression based on integer DTT is proposed.The compression efficiency of integer DTT is higher than integer DCT.The proposed lossless image compression scheme is compatible with JPEG.Both lossy and lossless image compression can be realized under the proposed scheme.A lossless color transform method based on traditional RGB to YCbCr is presented.

In spite of the impressive diversity of multi-factors fuzzy time series models, there is still a burning need to develop models that can handle the uncertainty inherent in certain data with some certain methods and obtain high forecasting accuracy. This paper proposes a novel multi-granularity combined prediction model for multi-factors fuzzy time series forecasting. The proposed model utilizes the clustering algorithm to generate different lengths of intervals, and forecasts the fuzzy trend in different granular spaces. It calculates the final forecasted value by using the forecasted fuzzy trend in each granular space and the optimal weighting vector obtained by particle swarm techniques. The proposed model can transform the uncertain problem to a certain problem through the granular computing theory. The experiments show that the presented model can not only obtain higher forecasting accuracy than several existing methods to forecast the Taiwan Stock Exchange Capitalization Weighted Stock Index (TAIEX) and the civilian unemployment rate, but also capture and interpret the fuzzy trend.