Electronic Letters on Computer Vision and Image Analysis (ELCVIA - Universitat Autònoma de Barcelona)
Not a member yet
343 research outputs found
Sort by
Behaviour understanding through the analysis of image sequences collected by wearable cameras
No full textDescribing people\u27s lifestyle has become a hot topic in the field of artificial intelligence. Lifelogging is described as the process of collecting personal activity data describing the daily behaviour of a person. Nowadays, the development of new technologies and the increasing use of wearable sensors allow to automatically record data from our daily living. In this paper, we describe our developed automatic tools for the analysis of collected visual data that describes the daily behaviour of a person. For this analysis, we rely on sequences of images collected by wearable cameras, which are called egocentric photo-streams. These images are a rich source of information about the behaviour of the camera wearer since they show an objective and first-person view of his or her lifestyle
Retinal Blood Vessel Extraction from Fundus Images Using Enhancement Filtering and Clustering
Full text linkScreening of vision troubling eye diseases by segmenting fundus images eases the danger of loss of sight of people. Computer assisted analysis can play an important role in the forthcoming health care system universally. Therefore, this paper presents a clustering based method for extraction of retinal vasculature from ophthalmoscope images. The method starts with image enhancement by contrast limited adaptive histogram equalization (CLAHE) from which feature extraction is accomplished using Gabor filter followed by enhancement of extracted features with Hessian based enhancement filters. It then extracts the vessels using K-mean clustering technique. Finally, the method ends with the application of a morphological cleaning operation to get the ultimate vessel segmented image. The performance of the proposed method is evaluated by taking two different publicly available Digital retinal images for vessel extraction (DRIVE) and Child heart and health study in England (CHASE_DB1) databases using nine different performance matrices. It gives average accuracies of 0.952 and 0.951 for DRIVE and CHASE_DB1 databases, respectively.
Algorithm for Iris recognition based on contourlet Transform and Entropy
No full textThe iris is one of the most secure biometric information that is widely employed in authentication systems. In this paper we present a method for iris recognition based on the Contourlet Transform and Entropy which entails i) the detection and segmentation of the iris, ii) its normalization, iii) the application of the Contourlet Transform, iv) the generation of the iris descriptor, and v) the matching between the query iris and those in the database. The proposed method has been tested with images taken from the popular CASIA-V4 and UBIRIS.v1 datasets and compared against four other iris recognition algorithms. The results show a higher true positive rate with a reduced computation time
Development of transition region based methods for image segmentation
Full text linkIn this thesis, some transition region based segmentation approaches have developed to perform image segmentation for grayscale and colour images.In computer vision and image understanding applications, image segmentation is an important pre-processing step. The main goal of the segmentation process is the separation of foreground region from background region. Based on the output of the segmentation result, segmentation can be categorized as global segmentation or local segmentation. The global segmentation aims for complete separation of the object from the background while the local segmentation divides the image into constituent regions. For achieving segmentation, a number of algorithms are developed by various researchers. The segmentation approaches are application specific and do not work well for both grayscale and colour image segmentation. For any image consisting of foreground and background, some transition regions exist between the foreground and background regions. Effective extraction of transition region leads to a better segmentation result. Therefore, the doctoral thesis intends to efficient and effective transition region approaches for image segmentation for both grayscale and colour images.The performance of the segmentation is qualitatively measured visually by looking at the ground truths as well as the segmentation masks generated from different segmentation approaches or by comparing the original image and the segmented result. The quantitative performance of segmentation results is compared via five performance measures as misclassification error (ME), false positive rate (FPR), false negative rate (FNR), Jaccard index (JI) and segmentation accuracy (SA).The doctoral research work is focused on the development of transition region approaches both in spatial and wavelet domain for image segmentation. The algorithms developed are categorized as(A) Grayscale transition region based segmentation approaches(B) Colour image transition region based segmentation approachesThe details are summarized below.(A) Grayscale transition region based segmentation approaches: Five transition region based approaches are developed for grayscale image segmentation (i) Proposed method 1, (ii) Proposed method 2, (iii) Proposed method 3, (iv) Proposed method 4 and (iv) Proposed method 5.(i) Proposed method 1: The Proposed method 1 extracts transition region using local variance and global thresholding considered from the variance features. The method utilizes edge linking and morphological operations for object extraction. The method is intended for segmentation of single and multiple objects from the image. The method does not perform well when the object and background gray level intensities are overlapping in nature.(ii) Proposed method 2: The Proposed method 2 utilizes 2-dimensional Gabor filter and global thresholding considered from the Gabor features for transition region extraction. Further, it uses edge linking and morphological operations for object extraction. The method is developed for eradicating the drawbacks of Proposed method 1. The performance of this method degrades in the presence of background texture.(iii) Proposed method 3: The Proposed method 3 is a hybridization of Proposed method 1 and Proposed method 2 in the transform domain. For suppressing the background textures, the wavelet transform is utilized in Proposed method 3. Utilizing the transitional features extracted by applying Proposed method 1 and Proposed method 2, a wavelet feature image is generated. The transition region is extracted by applying Otsu thresholding on these feature image. Further, edge linking and morphological operations are applied on it to extract the objects.(iv) Proposed method 4: The Proposed method 4 also utilizes the wavelet transform along with uses local standard deviation and Otsu thresholding for transition region extraction. The method suppresses the texture and extracts an effectively continuous transition region achieving better segmentation result. The method does not utilize the edge linking operation as used in Proposed method 1, Proposed method 2 and Proposed method 3. It uses less number of morphological operations as compared to that of Proposed method 1 and Proposed method 2.(v) Proposed method 5: The Proposed method uses fuzzy c-means clustering of local variance features along with Otsu thresholding for transition region extraction.(B) Colour image transition region based segmentation approaches: Here three transition region based colour image segmentation which is the extension of gray level transition region approaches are developed. The segmentation methods developed are (i) Proposed method 6, (ii) Proposed method 7 and (iii) Proposed method 8.(i) Proposed method 6: The Proposed method 6 is an extension of Proposed method 1 developed in RGB colour space.(ii) Proposed method 7: The Proposed method 7 is an extension of Proposed method 4 processed in CIE-Lab colour space. The method is applied to one agricultural application i.e. worm separation from infected plant leaves.(iii) Proposed method 8: The Proposed method 8 is a hybrid two-stage segmentation process which uses the algorithm of Proposed method 1 and Proposed method 2. This method processes the colour image in two different colour models (RGB model and CIE-Lab) for object extraction. The Proposed method 8 is applied for underwater fish image segmentation.The proposed methods are found to be better for segmentation of gray and colour images
Vostok: 3D scanner simulation for industrial robot environments
No full textComputer vision will drive the next wave of robot applications. Latest three-dimensional scanners provide increasingly realistic object reconstructions. We introduce an innovative simulator that allows interacting with those scanners within the operating environment, thus creating a powerful tool for developers, researchers and students. In particular, we present a novel approach for simulating structured-light and timeof-flight sensors. Qualitative results demonstrate the efficiency and reliability in industrial environments. By using the programmability of modern GPUs, it is now possible to make greater use of parallelized simulative approaches. Apart from the easy modification of sensor parameters, the main advantage in simulation is the opportunity of carrying out experiments under reproducible conditions, especially for dynamic scene setups. Moreover, thanks to a great computational power, it is possible to generate huge amounts of synthetic data which can be used as test datasets for training machine learning models
Automatic reactivity characterisation of char particles from pulverised coal combustion using computer vision
Full text linkChar morphologies produced during pulverised coal combustion may determine coal reactivity which affects the combustion efficiency and the emissions of CO2 in power plants. Commonly, char samples are characterised manually, but this process is subjective and time-consuming. This work proposes methods to automate the char reactivity characterisation using microscopy images and computer vision techniques. These methods are summarised in three contributions: the localisation of char particles based on candidate regions and deep learning methods; the classification of particles into char reactivity groups using morphological and texture features; and the integration in a system of the two previous proposals to characterise char sample reactivity. The proposed system successfully estimate char reactivity in a fast and accurate way
Design and Development of a Computer Vision Algorithm and Tool for Currency Recognition in Indian Vernacular Languages for Visually Challenged People
Full text linkThe God created this universe with all living and non-living entities. Human is one of the best among His creations. For human beings, eyes are the best gift of the God to see all His creations. As of now, human beings are considered as the only developed creatures among the God’s creations and have developed themselves from Stone Age to the Super Computing Era. As the human civilizations grew up, the day-to-day transactions moved from the barter system to the currency, the banknotes. Today, every country has its currency in terms of coins and paper notes. Each of the currency of individual country has its unique features, colors, denominations and international value. The life moves on this currency only. We, all, having been given two beautiful eyes could recognize the currency easily, but the same is not easy for the blind people. Though the denomination of a currency can easily be recognized to differentiate between counterfeit currencies from the real one is a Holy Grail. Especially for the blind people, it is a herculean task like finding a needle from a haystack. Since money is the cause of any cheating, if the person is blind, the chances of him being cheated are more. There are many tools available all over the world for the currencies of other developed countries. But, in India, there are no specific robust and handy tools that can help the blind people to recognize the Indian currencies in their mother tongue. For that reason, the main motive of this work is to develop and test a robust computer vision algorithm(s) to identify the Indian currency, mainly paper-based currency, in Indian regional languages.To go ahead with this research, along with the other image matching techniques, the ORB (Oriented FAST Rotated BRIEF) has been used as a feature detector. The reason behind the use of the ORB is the trade-off in the performance of the ORB. In its category, the ORB has been proved less accurate than its siblings the SIFT (Scale-Invariant Feature Transform) and the SURF (Speeded-Up Robust Features) in terms of feature detection and hence accuracy. However, the ORB is faster in terms of execution time than the others. As the SIFT and the SURF are patented technologies and ORB is the free and open source, this work attempts to improve the performance of the ORB in terms of recognition accuracy. In this direction, first, for preprocessing, the time performance of GrabCut algorithm has been improved (An algorithm which is used to remove the background from the images) for Android-based devices, named as cGrab-Cut. The output of this algorithm can be used for further processing of the image. For feature detection, two hybrid approaches have been developed to improve the performance of the ORB, named as HORB – A Histogram based ORB and ACORB – An ACO based ORB.In order to provide the best performance for image classification, this work lastly proposed, developed and tested two classifiers: a three-stage hybrid classifier, the HORBoVF which is based on the Histograms, ORB and Bag of Visual Words, and a two-stage ACO, ORB, and Bag of Visual Words based classifier, the ACORBoVF. All the proposed algorithms have been developed in such a way so that they can work in constrained environments like low memory and slow processors as well as for any images, not limited to Indian currencies only. Along with the two classifiers, a Transfer learning based image classifier, Te₹₹ency (Tensor Currency), has also been trained and tested to see the performance of convolutional neural network using TensorFlow technology and compare it with the two proposed classifiers. To test the effectiveness of all the proposed approaches, Python and Android-based test programs and tools have been developed. The results prove that all the proposed approaches serve the aforementioned motive and are much better as compared to the sole ORB. Thus, taking the advantage of faster execution of the ORB, this work tried to improve the performance of the ORB with different approaches for correct image identification.Apart from detecting the denominations, a histogram based analysis that has also been carried out. It gives a promising inference that the chemical composition of the printing colors used in currency notes can also be used to differentiate the real currencies from the fake one. For a country like India, where fake currencies are being dumped by black money hoarders and enemy countries, this technique would be proven useful for individuals who cannot have access to the bank-like verification machines to check the genuineness of the currencies. The technique has been tested on mobile phones to ensure that it is handy and can be included in App so that it can be affordable and accessible to everyone. Though, the output could be affected if the distance varies for image capture but this technique gives a new direction to move ahead for addressing the issue of fake currency detection
Automated Leaf Alignment and Partial Shape Feature Extraction for Plant Leaf Classification
Full text linkThe last few decades have witnessed various approaches to automate the process of plant classification using the characteristics of the leaf. Several approaches have been proposed, and the majority focused on global shape features. However, one challenge that faces this task is the high interclass similarity amongst the leaves of different species in terms of the global shape. Furthermore, there always has been an obstacle against full automation as several approaches require user intervention to align the leaf. Therefore, a new set of Quartile Features (QF) is proposed in this paper to describe the partial shape of the leaf, in addition to an automated alignment approach to automate the system. The QF are extracted from the horizontal and vertical leaf quartiles to describe the partial shape of the leaf and the relations among its parts. The well-known Flavia dataset has been selected for the evaluation of the proposed system. The experimental results indicate the ability of the proposed alignment algorithm to align leaves with different shapes and maintain a correct classification accuracy regardless of the orientation of the input leaf samples. Furthermore, the proposed QF indicated promising results by increasing the accuracy of the classification by a range of approximately 26% to 30% when combined with Hu’s Moment Invariants, using k-fold cross-validation technique
An Adaptive Non-linear Statistical Salt-and-Pepper Noise Removal Algorithm using Interquartile Range
Full text linkThis paper presents a salt-and-pepper noise removal scheme using modified mean filter. The proposed method is based on a simple basic concepts of mean filter, where each mean value is calculated from the mathematical formula of interquartile range (IQR). It replaces the noisy pixels using IQR based mathematical formula applied on the filter window. Experimental results are presented to demonstrate the efficiency (quality of the image) of the method compared to other existing different types of impulse noise removal techniques
Image decomposition using a second-order variational model and wavelet shrinkage
Full text linkThe paper is devoted to the new model for image decomposition, that splits an image into three components , with a piecewise-smooth or the ``cartoon\u27\u27 component, a texture component and the noise part in variational approach. This decomposition model is in fact incorporates the advantages of two preceding models: the second-order total variation minimization of Rudin-Osher-Fatemi (ROF2), and wavelet shrinkage for oscillatory functions. This decomposition model is presented as an extension of the three components decomposition algorithm of Aujol et al. in \cite{JAC}. It also continues the idea introduced previously by authors in \cite{TPB}, for two components decomposition model. The ROF2 model was first proposed by Bergounioux et al. in \cite{BP}, it is an improved regularization method to overcome the undesirable staircasing effect. The wavelet shrinkage is well combined to separate the oscillating part due to texture from that due to noise. Experimental results validate the proposed algorithm and demonstrate that the image decomposition model presents effective and comparable performance to other state-of-the-art models