Vehicle Matching Based on Similarity Metric Learning

With the development of new media technology,vehicle matching plays a further significant role in video surveillance systems.Recent methods explored the vehicle matching based on the feature extraction.Meanwhile,similarity metric learning also has achieved enormous progress in vehicle matching.But most of these methods are less effective in some realistic scenarios where vehicles usually be captured in different times.To address this cross-domain problem,we propose a cross-domain similarity metric learning method that utilizes theGANto generate vehicle imageswith another domain and propose the two-channel Siamese network to learn a similarity metric from both domains(i.e.,Day pattern or Night pattern)for vehicle matching.To exploit properties and relationships among vehicle datasets,we first apply the domain transformer to translate the domain of vehicle images,and then utilize the two-channel Siamese network to extract features from both domains for better feature similarity learning.Experimental results illustrate that our models achieve improvements over state-of-the-arts.

Learning to combine multiple string similarity metrics for effective toponym matching

Several tasks related to geographical information retrieval and to the geographical information sciences involve toponym matching,that is,the problem of matching place names that share a common referent.In this article,we present the results of a wide-ranging evaluation on the performance of different string similarity metrics over the toponym matching task.We also report on experiments involving the usage of supervised machine learning for combining multiple similarity metrics,which has the natural advantage of avoiding the manual tuning of similarity thresholds.Experiments with a very large dataset show that the performance differences for the individual similarity metrics are relatively small,and that carefully tuning the similarity threshold is important for achieving good results.The methods based on supervised machine learning,particularly when considering ensembles of decision trees,can achieve good results on this task,significantly outperforming the individual similarity metrics.

Rock mass structural recognition from drill monitoring technology in underground mining using discontinuity index and machine learning techniques

A procedure to recognize individual discontinuities in rock mass from measurement while drilling(MWD)technology is developed,using the binary pattern of structural rock characteristics obtained from in-hole images for calibration.Data from two underground operations with different drilling technology and different rock mass characteristics are considered,which generalizes the application of the methodology to different sites and ensures the full operational integration of MWD data analysis.Two approaches are followed for site-specific structural model building:a discontinuity index(DI)built from variations in MWD parameters,and a machine learning(ML)classifier as function of the drilling parameters and their variability.The prediction ability of the models is quantitatively assessed as the rate of recognition of discontinuities observed in borehole logs.Differences between the parameters involved in the models for each site,and differences in their weights,highlight the site-dependence of the resulting models.The ML approach offers better performance than the classical DI,with recognition rates in the range 89%to 96%.However,the simpler DI still yields fairly accurate results,with recognition rates 70%to 90%.These results validate the adaptive MWD-based methodology as an engineering solution to predict rock structural condition in underground mining operations.

Robust multi-sensor image matching based on normalized self-similarity region descriptor

Multi-modal image matching is crucial in aerospace applications because it can fully exploit the complementary and valuable information contained in the amount and diversity of remote sensing images.However,it remains a challenging task due to significant non-linear radiometric,geometric differences,and noise across different sensors.To improve the performance of heterologous image matching,this paper proposes a normalized self-similarity region descriptor to extract consistent structural information.We first construct the pointwise self-similarity region descriptor based on the Euclidean distance between adjacent image blocks to reflect the structural properties of multi-modal images.Then,a linear normalization approach is used to form Modality Independent Region Descriptor(MIRD),which can effectively distinguish structural features such as points,lines,corners,and flat between multi-modal images.To further improve the matching accuracy,the included angle cosine similarity metric is adopted to exploit the directional vector information of multi-dimensional feature descriptors.The experimental results show that the proposed MIRD has better matching accuracy and robustness for various multi-modal image matching than the state-of-the-art methods.MIRD can effectively extract consistent geometric structure features and suppress the influence of SAR speckle noise using non-local neighboring image blocks operation,effectively applied to various multi-modal image matching.

Improved Collaborative Filtering Recommendation Based on Classification and User Trust

When dealing with the ratings from users,traditional collaborative filtering algorithms do not consider the credibility of rating data,which affects the accuracy of similarity.To address this issue,the paper proposes an improved algorithm based on classification and user trust.It firstly classifies all the ratings by the categories of items.And then,for each category,it evaluates the trustworthy degree of each user on the category and imposes the degree on the ratings of the user.Finally,the algorithm explores the similarities between users,finds the nearest neighbors,and makes recommendations within each category.Simulations show that the improved algorithm outperforms the traditional collaborative filtering algorithms and enhances the accuracy of recommendation.

An algorithm for 3D shape matching using spherical sectioning

3D shape searching is a problem of current interest in several different fields. Most techniques are developed for a particular domain and used to reduce a shape into a simpler shape representation. The techniques developed for a particular domain will also find application in other domains. We propose a new shape matching method. The SSRD (spherical sectioning railroad diagram) algorithm has the general shape distribution’s properties and overall features of the original model. The SSRD’s useful properties are discussed. We show the experimental results for the validity of our method.

Fault Diagnosis of Power Transformers Using Graph Convolutional Network

Existing methods for transformer fault diagnosis either train a classifier to fit the relationship between dissolved gas and fault type or find some similar cases with unknown samples by calculating the similarity metrics.Their accuracy is limited,since they are hard to learn from other algorithms to improve their own performance.To improve the accuracy of transformer fault diagnosis,a novel method for transformer fault diagnosis based on graph convolutional network(GCN)is proposed.The proposed method has the advantages of two kinds of existing methods.Specifically,the adjacency matrix of GCN is utilized to fully represent the similarity metrics between unknown samples and labeled samples.Furthermore,the graph convolutional layers with strong feature extraction ability are used as a classifier to find the complex nonlinear relationship between dissolved gas and fault type.The back propagation algorithm is used to complete the training process of GCN.The simulation results show that the performance of GCN is better than that of the existing methods such as convolutional neural network,multi-layer perceptron,support vector machine,extreme gradient boosting tree,k-nearest neighbors and Siamese network in different input features and data volumes,which can effectively meet the needs of diagnostic accuracy.

A combination of learning and non-learning based method for enhancement, compression and reconstruction of underwater images

Complex nature of underwater environment poses biggest challenge towards image acquisition and transmission of underwater images.This paper proposes an integrated approach which consists of a non-learning enhancement method with deep Convolutional Neural Networks(CNN)for compression and reconstruction of the image.The proposed method does color and contrast correction for image enhancement.The enhanced images are down-sampled using 9-layer CNN followed by Discrete Wavelet Transform(DWT).The decompression is done by using Inverse DWT.Further,the sub-pixel up-sampled image is de-blurred using a three-layer CNN.Residual Dense CNN(RD-CNN)is used to improve the quality of the reconstructed image after deblurring.The quality of the reconstructed images is measured using Peak Signal to Noise Ratio(PSNR)and Structural Similarity Index Metric(SSIM).The proposed model provides better image enhancement,compression,and reconstruction quality than the existing state-of-the-art methods and Super Resolution CNN(SRCNN)respectively.