Search

With the continuous development of deep learning, its drawbacks are also beginning to appear. As an alternative to deep learning, broad learning is emerging. However, the level of broad learning model is shallow, so feature learning is not sufficient. In order to solve two problems of small samples whose dimensions are not very high in network model training, it cannot be adequately trained in the deep learning model and the features of input data cannot be fully learned in the broad learning model. This paper attempts to add a hidden layer on the enhancement nodes of the broad learning model and do shallow learning for the enhanced features to learn the hierarchical features again. This improved broad learning model also provides a new idea for solving the problem of small samples. From the perspective of regression and classification, this paper proves that for small samples whose dimensions are not very high, the effect of the improved broad learning model is better than that of the original broad learning model and also proves that the improved broad learning model has a good ability of application. This shows that the broad learning model based on enhanced features learning has the necessity and feasibility of further research.

In this paper, we propose a novel quality assessment of finger-vein images for quality control in the enrollment and authentication of a finger-vein verification system. First, a Radon transform based model is employed to assess the quality of a finger-vein grayscale image. Second, to assess the quality of a finger-vein binary image, we further proposed three evaluation functions to measure the connectivity, smoothness and reliability of the binary version of the finger-vein image. Finally, the scores from the finer-vein binary images are fused with these from finger-vein grayscale images to improve the performance. Experimental results show that our approach can effectively identify the low quality finger-vein images, which is also helpful in improving the performance of the finger-vein verification system. We also show that instead of choosing the images with the highest quality as the enrollment templates, using the templates with the mid-range quality would achieve better performance with respect to improvement of varication accuracy.

In order to strengthen the management of sand excavation in river courses and prevent the occurrence of illegal sand mining activities, this paper proposes an approach for monitoring sand mining based on sound. Firstly, Mel Frequency Cepstral Coefficients (MFCCs) abstractor and Autoencoder are combined to extract features of every frame of a sound sequence, and then each frame of the sound sequence is classified by a specific classifier. Finally a voting strategy is used among the frames to determine the final category of the sound sequence. Experiments show that whether the classifier is SVM, KNN, or BP neural network, the result of combined features is better than the result of features extracted by the MFCC abstractor. Therefore, it is feasible to use the artificial intelligence method based on sound features extracted through MFCC abstractor and Autoencoder to monitor sand mining.

This paper presents a new scheme to improve the performance of finger-vein identification systems. Firstly, a vein pattern extraction method to extract the finger-vein shape and orientation features is proposed. Secondly, to accommodate the potential local and global variations at the same time, a region-based matching scheme is investigated by employing the Scale Invariant Feature Transform (SIFT) matching method. Finally, the finger-vein shape, orientation and SIFT features are combined to further enhance the performance. The experimental results on databases of 426 and 170 fingers demonstrate the consistent superiority of the proposed approach.

This paper is aimed at the image encryption scheme applicable to JPEG2000 codec. Firstly, two chaotic maps are suggested and their statistic characteristics are also analyzed. Secondly, to accomplish a controllable visual effect of encrypted image, a visual quality control model is presented, and on the basis of which, a chaotic image encryption scheme is constructed by chaotically modulating the randomly selected approximate coefficients at the coarsest level in wavelet domain. Then, the security and efficiency of the algorithm are analyzed. Finally, the experiment results illustrate that the proposed algorithms are credible, secure, efficient, and practical for JPSEC.

Based on the characteristic of dyadic wavelet transform to image denoising , this paper presents that denoising precision can be improved by the way that adopting different thresholds according to the different scales of the wavelet coefficients of image and noise to establish the self-adjusted layered threshold function which adapts to it and reconstruct the wavelet. The experiment shows that by the method above the effect of image denoising is obviously superior to that of fixed threshold.

Recent work has shown improvements in text clustering and classification by integrating conceptual features extracted from background knowledge. In this paper we address the problem of text classification with labeled data and unlabeled data. We propose a Latent Bayes Ensemble model based on word-concept mapping and transductive boosting method. With the knowledge extracted from ontologies, we hope to improve the classification accuracy even with large amounts of unlabeled documents. We conducted several experiments on two well-known corpora and the results are compared with Naive Bayes and TSVM classifiers.