We analyze the drawbacks of DBSCAN and its variants, and find the grid technique, which is used in Fast-DBSCAN and ρ-approximate DBSCAN, is almost useless in high dimensional data space. Because it ...usually yields considerable redundant distance computations. In order to tame these problems, two techniques are proposed: one is to use ϵ2-norm ball to identify Inner Core Blocks within which all points are core points, it has higher efficiency than grid technique for finding more core points at one time; the other is a fast approximate algorithm for judging whether two Inner Core Blocks are density-reachable from each other. Besides, cover tree is also used to accelerate the process of density computations. Based on the three techniques, an approximate approach, namely BLOCK-DBSCAN, is proposed for large scale data, which runs in about O(nlog (n)) expected time and obtains almost the same result as DBSCAN. BLOCK-DBSCAN has two versions, i.e., L2 version can work well for relatively high dimensional data, and L∞ version is suitable for high dimensional data. Experimental results show that BLOCK-DBSCAN is promising and outperforms NQDBSCAN, ρ-approximate DBSCAN and AnyDBC.
•The underlying idea is: point p and point q should have similar neighbors, provided p and q are close to each other; given a certain eps, the closer they are, the more similar their neighbors ...are.•NQ-DBSCAN is an exact algorithm that may return the same result as DBSCAN if the parameters are same. While ρ-Approximate DBSCAN is an approximate algorithm.•The best complexity of NQ-DBSCAN can be O(n), and the average complexity of NQ-DBSCAN is proved to be O(n log(n)) provided the parameters are properly chosen. While ρ-Approximate DBSCAN runs only in O(n2) in high dimension.•NQ-DBSCAN is suitable for clustering data with a lot of noise.
Clustering is an important technique to deal with large scale data which are explosively created in internet. Most data are high-dimensional with a lot of noise, which brings great challenges to retrieval, classification and understanding. No current existing approach is “optimal” for large scale data. For example, DBSCAN requires O(n2) time, Fast-DBSCAN only works well in 2 dimensions, and ρ-Approximate DBSCAN runs in O(n) expected time which needs dimension D to be a relative small constant for the linear running time to hold. However, we prove theoretically and experimentally that ρ-Approximate DBSCAN degenerates to an O(n2) algorithm in very high dimension such that 2D > > n. In this paper, we propose a novel local neighborhood searching technique, and apply it to improve DBSCAN, named as NQ-DBSCAN, such that a large number of unnecessary distance computations can be effectively reduced. Theoretical analysis and experimental results show that NQ-DBSCAN averagely runs in O(n*log(n)) with the help of indexing technique, and the best case is O(n) if proper parameters are used, which makes it suitable for many realtime data.
This paper proposes an end-to-end temporal attention learning method to improve the performance of action quality assessment in sports video. For temporal weighted training, an attention-learning ...module is built to simulate the attention mechanism and judgement preference of human perception on action quality assessment. The weights are learned based on the loss of the segmented prediction errors and used to balance the significance of segmented features. We evaluate the proposed method on diving and gym-vault action of the benchmark AQA-7 dataset. The experimental results show that the proposed attention-aware feature training method is more effective than temporal aggregation and existing temporal relationship learning methods. Furthermore, only using the distance loss between the predicated score and the ground-truth score, without considering the ranking loss of different videos for training, this paper has achieved the state-of-the-art performance on both of the spearman rank correlation and mean Euclidean distance of the predicted scores against the judge’s scores.
Automatic road extraction from remote sensing images plays an important role for navigation, intelligent transportation, and road network update, etc. Convolutional neural network (CNN)-based methods ...have presented many achievements for road extraction from remote sensing images. CNN-based methods require a large dataset with high quality labels for model training. However, there is still few standard and large dataset, which is specially designed for road extraction from optical remote sensing images. Besides, the existing end-to-end CNN models for road extraction from remote sensing images are usually with symmetric structure, studying on asymmetric structure between encoding and decoding is rare. To address the above problems, this article first provides a publicly available dataset LRSNY for road extraction from optical remote sensing images with manually labelled labels. Second, we propose a reconstruction bias U-Net for road extraction from remote sensing images. In our model, we increase the decoding branches to obtain multiple semantic information from different upsamplings. Experimental results show that our method achieves better performance compared with other six state-of-the-art segmentation models when testing on our LRSNY dataset. We also test on Massachusetts and Shaoshan datasets. The good performances on the two datasets further prove the effectiveness of our method.
Lipreading with DenseNet and resBi-LSTM Chen, Xuejuan; Du, Jixiang; Zhang, Hongbo
Signal, image and video processing,
07/2020, Letnik:
14, Številka:
5
Journal Article
Recenzirano
Lipreading is to recognize what the speakers say by the movement of lip only. Most of the previous works are to solve the problem of lipreading in English. For Mandarin lipreading, there are a few ...researches due to the lack of datasets. For that reason, we introduce a simple method here to build a dataset for sentence-level Mandarin lipreading from programs like news, speech and talk show. We use Hanyu Pinyin (a phonemic transcription of Chinese) as label and totally have 349 classes, while the number of Chinese characters is 1705 in our dataset. Therefore, for lipreading, there are two steps. The first step is to obtain the Hanyu Pinyin sequence. We propose a model that is composed of a 3D convolutional layer with DenseNet and residual bidirectional long short-term memory. After this, in order to get the final Chinese characters results, a model with a stack of multi-head attention is applied to convert Hanyu Pinyin into Chinese characters.
•We propose an Adaboost-like combination strategy of multiple U-Nets.•We design multiple U-Nets’ combination as multiple-objective optimization problem, thus we jointly train multiple U-Nets.•Our ...model can achieve quite good performance with much less parameters compared with several state-of-the-art models.•We verify the performance of our model on three road datasets: LRSNY, Shaoshan and Massachusetts datasets.
Road extraction from optical remote sensing images has many important application scenarios, such as navigation, automatic driving and road network planning, etc. Current deep learning based models have achieved great successes in road extraction. Most deep learning models improve abilities rely on using deeper layers, resulting to the obese of the trained model. Besides, the training of a deep model is also difficult, and may be easy to fall into over fitting. Thus, this paper studies to improve the performance through combining multiple lightweight models. However, in fact multiple isolated lightweight models may perform worse than a deeper and larger model. The reason is that those models are trained isolated. To solve the above problem, we propose an Adaboost-like End-To-End Multiple Lightweight U-Nets model (AEML U-Nets) for road extraction. Our model consists of multiple lightweight U-Net parts. Each output of prior U-Net is as the input of next U-Net. We design our model as multiple-objective optimization problem to jointly train all the U-Nets. The approach is tested on two open datasets (LRSNY and Massachusetts) and Shaoshan dataset. Experimental results prove that our model has better performance compared with other state-of-the-art semantic segmentation methods.
Benefiting from big data, powerful computation and new algorithmic techniques, we have been witnessing the renaissance of deep learning, particularly the combination of natural language processing ...(NLP) and deep neural networks. The advent of electronic medical records (EMRs) has not only changed the format of medical records but also helped users to obtain information faster. However, there are many challenges regarding researching directly using Chinese EMRs, such as low quality, huge quantity, imbalance, semi-structure and non-structure, particularly the high density of the Chinese language compared with English. Therefore, effective word segmentation, word representation and model architecture are the core technologies in the literature on Chinese EMRs.
In this paper, we propose a deep learning framework to study intelligent diagnosis using Chinese EMR data, which incorporates a convolutional neural network (CNN) into an EMR classification application. The novelty of this paper is reflected in the following: (1) We construct a pediatric medical dictionary based on Chinese EMRs. (2) Word2vec adopted in word embedding is used to achieve the semantic description of the content of Chinese EMRs. (3) A fine-tuning CNN model is constructed to feed the pediatric diagnosis with Chinese EMR data. Our results on real-world pediatric Chinese EMRs demonstrate that the average accuracy and F1-score of the CNN models are up to 81%, which indicates the effectiveness of the CNN model for the classification of EMRs. Particularly, a fine-tuning one-layer CNN performs best among all CNNs, recurrent neural network (RNN) (long short-term memory, gated recurrent unit) and CNN-RNN models, and the average accuracy and F1-score are both up to 83%.
The CNN framework that includes word segmentation, word embedding and model training can serve as an intelligent auxiliary diagnosis tool for pediatricians. Particularly, a fine-tuning one-layer CNN performs well, which indicates that word order does not appear to have a useful effect on our Chinese EMRs.
Human activity analysis in the legal monitoring environment plays an important role in the physical rehabilitation field, as it helps patients with physical injuries improve their postoperative ...conditions and reduce their medical costs. Recently, several deep learning-based action quality assessment (AQA) frameworks have been proposed to evaluate physical rehabilitation exercises. However, most of them treat this problem as a simple regression task, which requires both the action instance and its score label as input. This approach is limited by the fact that the annotations in this field usually consist of healthy or unhealthy labels rather than quality scores provided by professional physicians. Additionally, most of these methods cannot provide informative feedback on a patient's motion defects, which weakens their practical application. To address these problems, we propose a multi-task contrastive learning framework to learn subtle and critical differences from skeleton sequences to deal with the performance metric and AQA problems of physical rehabilitation exercises. Specifically, we propose a performance metric network that takes triplets of training samples as input for score generation. For the AQA task, the same contrast learning strategy is used, but pairwise training samples are fed into the action quality assessment network for score prediction. Notably, we propose quantifying the deviation of the joint attention matrix between different skeleton sequences and introducing it into the loss function of our learning network. It is proven that considering both score prediction loss and joint attention deviation loss improves physical exercises AQA performance. Furthermore, it helps to obtain informative feedback for patients to improve their motion defects by visualizing the joint attention matrix's difference. The proposed method is verified on the UI-PRMD and KIMORE datasets. Experimental results show that the proposed method achieves state-of-the-art performance.
Traumatic brain injury (TBI), known as intracranial injury, has been a serious threat to human health. Evidence exists indicating that autophagy and inflammatory responses contribute to secondary ...brain injury after TBI. Notably, receptor-interacting protein kinase 1 (Ripk1) exerts an important role in cell autophagy. Therefore, this study aims to explore the effect of Ripk1 on neuron autophagy and apoptosis in TBI. Initially, blood samples of patients with TBI and healthy persons were collected to detect the expression of Ripk1, nuclear factor-kappa B (NF-κB), and NF-kB inhibitor α (IKBα). Then rat models with TBI were successfully established and, respectively, treated with shRNA targeting Ripk1 (sh-Ripk1), Ripk1 overexpression plasmid (oe-Ripk1), or IKKα inhibitor (BAY 11-7082). Subsequently, reverse transcription quantitative polymerase chain reaction and Western blot analysis were conducted to detect the expression of Ripk1, IKBα, NF-κB signaling pathway-, and apoptosis-related factors. Enzyme-linked immunosorbent assay was used to detect the expression of inflammatory cytokines. Compared with healthy persons, the expression of Ripk1, NF-κB and IKBα in blood of TBI patients was significantly upregulated. After silencing of Ripk1 or inhibition of the NF-κB signaling pathway, the expression of IL-1β, IL-6, TNF-α, Bax, and cleaved-caspase-3 was downregulated, and the expression of Bcl-2, ATG5, and LC3II/LC3I was upregulated. Furthermore, neuron injury and apoptosis were notably reduced and neuron autophagy increased significantly by Ripk1 downregulation or IKKα inhibitor. Ripk1 overexpression contributed to activation of NF-κB signaling pathway, whereby aggravating TBI-induced damage. Silencing Ripk1 suppresses TBI by inhibiting inflammation and promoting autophagy of neurons via inhibition of NF-κB signaling pathway.
Inverting a matrix is time-consuming, and many works focus on accelerating the inversion of a single large matrix by GPU. However, the problem of parallelizing the inversion of a large number of ...small matrices has received little attention. These problems are widely applied in computer science, including accelerating cryptographic algorithms and image processing algorithms. In this paper, we propose a Revised In-Place Inversion algorithm for inverting a large number of small matrices on the CUDA platform, which adopts a more refined parallelization scheme and outperforms other algorithms, achieving a speedup of up to 20.9572 times over the batch matrix inverse kernel in CUBLAS. Additionally, we found that there is an upper bound on the input data size for each GPU device, and the performance will degrade if the input data size is too large. Therefore, we propose the Saturation Size Curve based on this finding to divide matrices into batches and improve the algorithm performance. Experimental results show that this strategy increases the algorithm’s performance by 1.75 times and effectively alleviates the problem of performance degradation.
PDF
