Efficient Image Representation Research Articles

An Improved Retinal Blood Vessel Detection System Using an Extreme Learning Machine

Retinal images are commonly used to diagnose various diseases, such as diabetic retinopathy, glaucoma, and hypertension. An important step in the analysis of such images is the detection of blood vessels, which is usually done manually and is time consuming. The main proposal in this work is a fast method for retinal blood vessel detection using Extreme Learning Machine (ELM). ELM requires only one iteration to complete its training and it is a robust and fast network in all aspects. The proposal is a compact and efficient representation of retinal images in which the authors achieved a reduction up to 39% of the initial data volume, while still keeping representativeness. To achieve such a reduction whilst maintaining the representativeness, three features (local tophat, local average, and local variance) were used. According to the simulations carried out, this proposal achieved an accuracy of about 95% for most results, outperforming most of the state-of-art methods. Furthermore, this proposal has greater sensitivity, meaning that more vessel pixels are detected correctly.

Performance enhanced ripplet transform based compression method for medical images

Ripplet transforms provide efficient representation of images with all the spatial details. Since more details are extracted, the compression ratio is usually less. In this work, lossless prediction and lossy singular value decomposition of the ripplet coefficients are carried out to improve the performance of conventional ripplet transform based system. Initially, lossless prediction exploits the correlation between the image pixels and eases the process of transformation. Fast and computationally efficient randomized singular value decomposition technique is used to capture the essential information of the high frequency ripplet coefficients. The low frequency and the high frequency components are then encoded using entropy method. Experimental results show promising results for the proposed approached in terms of performance measure.

Supervised and Unsupervised End-to-End Deep Learning for Gene Ontology Classification of Neural In Situ Hybridization Images.

In recent years, large datasets of high-resolution mammalian neural images have become available, which has prompted active research on the analysis of gene expression data. Traditional image processing methods are typically applied for learning functional representations of genes, based on their expressions in these brain images. In this paper, we describe a novel end-to-end deep learning-based method for generating compact representations of in situ hybridization (ISH) images, which are invariant-to-translation. In contrast to traditional image processing methods, our method relies, instead, on deep convolutional denoising autoencoders (CDAE) for processing raw pixel inputs, and generating the desired compact image representations. We provide an in-depth description of our deep learning-based approach, and present extensive experimental results, demonstrating that representations extracted by CDAE can help learn features of functional gene ontology categories for their classification in a highly accurate manner. Our methods improve the previous state-of-the-art classification rate (Liscovitch, et al.) from an average AUC of 0.92 to 0.997, i.e., it achieves 96% reduction in error rate. Furthermore, the representation vectors generated due to our method are more compact in comparison to previous state-of-the-art methods, allowing for a more efficient high-level representation of images. These results are obtained with significantly downsampled images in comparison to the original high-resolution ones, further underscoring the robustness of our proposed method.

Adaptive particle representation of fluorescence microscopy images

Modern microscopes create a data deluge with gigabytes of data generated each second, and terabytes per day. Storing and processing this data is a severe bottleneck, not fully alleviated by data compression. We argue that this is because images are processed as grids of pixels. To address this, we propose a content-adaptive representation of fluorescence microscopy images, the Adaptive Particle Representation (APR). The APR replaces pixels with particles positioned according to image content. The APR overcomes storage bottlenecks, as data compression does, but additionally overcomes memory and processing bottlenecks. Using noisy 3D images, we show that the APR adaptively represents the content of an image while maintaining image quality and that it enables orders of magnitude benefits across a range of image processing tasks. The APR provides a simple and efficient content-aware representation of fluosrescence microscopy images.

Unifying Sum and Weighted Aggregations for Efficient yet Effective Image Representation Computation.

Embedding and aggregating a set of local descriptors (e.g. SIFT) into a single vector is normally used to represent images in image search. Standard aggregation operations include sum and weighted aggregations. While showing high efficiency, sum aggregation lacks discriminative power. In contrast, weighted aggregation shows promising retrieval performance but suffers extremely high time cost. In this work, we present a general mixed aggregation method that unifies sum and weighted aggregation methods. Owing to its general formulation, our method is able to balance the trade-off between retrieval quality and image representation efficiency. Additionally, to improve query performance, we propose computing multiple weighting coefficients rather than one for each to be aggregated vector by partitioning them into several components with negligible computational cost. Extensive experimental results on standard public image retrieval benchmarks demonstrate that our aggregation method achieves state-of-the-art performance while showing over ten times speedup over baselines.

Bundled Local Features for Image Representation

Local features have been widely used for image representation. Traditional methods often treat each local feature independently or simply model the correlations of local features with spatial partition. However, local features are correlated and should be jointly modeled. Besides, due to the variety of images, predefined partition rules will probably introduce noisy information. To solve these problems, in this paper we propose a novel bundled local features method for efficient image representation and apply it for classification. Specially, we first extract local features and bundle them together with over-complete spatial shapes by viewing each local feature as the central point. Then, the most discriminatively bundling features are selected by reconstruction error minimization. The encoding parameters are then used for image representations in a matrix form. Finally, we train bi-linear classifiers with quadratic hinge loss to predict the classes of images. The proposed method can combine local features appropriately and efficiently for discriminative representations. Experimental results on three image data sets show the effectiveness of the proposed method compared with other local features combination strategies.

Unifying Visual Attribute Learning with Object Recognition in a Multiplicative Framework.

Attributes are mid-level semantic properties of objects. Recent research has shown that visual attributes can benefit many typical learning problems in computer vision community. However, attribute learning is still a challenging problem as the attributes may not always be predictable directly from input images and the variation of visual attributes is sometimes large across categories. In this paper, we propose a unified multiplicative framework for attribute learning, which tackles the key problems. Specifically, images and category information are jointly projected into a shared feature space, where the latent factors are disentangled and multiplied to fulfil attribute prediction. The resulting attribute classifier is category-specific instead of being shared by all categories. Moreover, our model can leverage auxiliary data to enhance the predictive ability of attribute classifiers, which can reduce the effort of instance-level attribute annotation to some extent. By integrated into an existing deep learning framework, our model can both accurately predict attributes and learn efficient image representations. Experimental results show that our method achieves superior performance on both instance-level and category-level attribute prediction. For zero-shot learning based on visual attributes and human-object interaction recognition, our method can improve the state-of-the-art performance on several widely used datasets.

Analysis of large-scale UAV images using a multi-scale hierarchical representation

Unmanned aerial vehicle (UAV)-based imaging systems have many superiorities compared with other platforms, such as high flexibility and low cost in collecting images, providing wide application prospects. However, the acquisition of the UAV-based image commonly results in very high resolution and very large-scale images, which poses great challenges for subsequent applications. Therefore, an efficient representation of large-scale UAV images is necessary for the extraction of the required information in a reasonable time. In this work, we proposed a multi-scale hierarchical representation, i.e. binary partition tree, for analyzing large-scale UAV images. More precisely, we first obtained an initial partition of images by an oversegmentation algorithm, i.e. the simple linear iterative clustering. Next, we merged the similar superpixels to build an object-based hierarchical structure by fully considering the spectral and spatial information of the superpixels and their topological relationships. Moreover, objects of interest and optimal segmentation were obtained using object-based analysis methods with the hierarchical structure. Experimental results on processing the post-seismic UAV images of the 2013 Ya’an earthquake and the mosaic of images in the South-west of Munich demonstrate the effectiveness and efficiency of our proposed method.

An improved ORNAM representation of grey images

An efficient image representation can save space and facilitate the manipulation of the acquired images. In order to further enhance the reconstructed image quality and reduce the number of the homogeneous blocks of the overlapping rectangular non-symmetry and anti-packing model (ORNAM) representation, in this paper, we propose an improved overlapping rectangular non-symmetry and anti-packing model representation (IORNAM) of grey images. Compared with most of the up-to-date and the state-of-the-art hierarchical representation methods, the new IORNAM representation is characterised by two properties: 1) it adopts a ratio parameter of the length and the width of a homogenous block to improve the reconstructed image quality; 2) it uses a new expansion method to anti-pack the subpatterns of grey images to further decrease the number of homogenous blocks, which is important for improving the compression ratios of image representation and reducing the complexities of many image manipulation algorithms. The experimental results presented in this paper demonstrate that: 1) the new IORNAM representation is able to achieve high representation efficiency for grey images; 2) the new IORNAM representation outperforms most of the up-to-date and the state-of-the-art hierarchical representation methods of grey images.

A Multiscale Deeply Described Correlatons-Based Model for Land-Use Scene Classification

Research efforts in land-use scene classification is growing alongside the popular use of High-Resolution Satellite (HRS) images. The complex background and multiple land-cover classes or objects, however, make the classification tasks difficult and challenging. This article presents a Multiscale Deeply Described Correlatons (MDDC)-based algorithm which incorporates appearance and spatial information jointly at multiple scales for land-use scene classification to tackle these problems. Specifically, we introduce a convolutional neural network to learn and characterize the dense convolutional descriptors at different scales. The resulting multiscale descriptors are used to generate visual words by a general mapping strategy and produce multiscale correlograms of visual words. Then, an adaptive vector quantization of multiscale correlograms, termed multiscale correlatons, are applied to encode the spatial arrangement of visual words at different scales. Experiments with two publicly available land-use scene datasets demonstrate that our MDDC model is discriminative for efficient representation of land-use scene images, and achieves competitive classification results with state-of-the-art methods.

Embedded deep vision in smart cameras for multi-view objects representation and retrieval

Active large scale surveillance of indoor and outdoor environments with multiple cameras is becoming an undeniable necessity in today's connected world. Enhanced computational and storage capabilities in smart cameras establish them as promising platforms for implementing intelligent and autonomous surveillance networks. However, poor resolution, limited number of samples per object, and pose variation in multi-view surveillance streams, make the task of efficient image representation highly challenging. To address these issues, we propose an efficient and powerful convolutional neural network (CNN) based framework for features extraction using embedded processing on smart cameras. Efficient, high performance, pre-trained CNNs are separately fine-tuned on persons and vehicles to obtain discriminative, low dimensional features from segmented surveillance objects. Furthermore, multi-view queries of surveillance objects are used to improve retrieval performance. Experiments reveal better efficiency and retrieval performance in different surveillance datasets.

Face Representation and Face Recognition using Optimized Local Ternary Patterns (OLTP)

For many years, researchers in face description area have been representing and recognizing faces based on different methods that include subspace discriminant analysis, statistical learning and non-statistics based approach etc. But still automatic face recognition remains an interesting but challenging problem. This paper presents a novel and efficient face image representation method based on Optimized Local Ternary Pattern (OLTP) texture features. The face image is divided into several regions from which the OLTP texture feature distributions are extracted and concatenated into a feature vector that can act as face descriptor. The recognition is performed using nearest neighbor classification method with Chi-square distance as a similarity measure. Extensive experimental results on Yale B, ORL and AR face databases show that OLTP consistently performs much better than other well recognized texture models for face recognition.

A ganglion-cell-based primary image representation method and its contribution to object recognition

ABSTRACTA visual stimulus is represented by the biological visual system at several levels: in the order from low to high levels, they are: photoreceptor cells, ganglion cells (GCs), lateral geniculate nucleus cells and visual cortical neurons. Retinal GCs at the early level need to represent raw data only once, but meet a wide number of diverse requests from different vision-based tasks. This means the information representation at this level is general and not task-specific. Neurobiological findings have attributed this universal adaptation to GCs' receptive field (RF) mechanisms. For the purposes of developing a highly efficient image representation method that can facilitate information processing and interpretation at later stages, here we design a computational model to simulate the GC's non-classical RF. This new image presentation method can extract major structural features from raw data, and is consistent with other statistical measures of the image. Based on the new representation, the performances of other state-of-the-art algorithms in contour detection and segmentation can be upgraded remarkably. This work concludes that applying sophisticated representation schema at early state is an efficient and promising strategy in visual information processing.

Efficient tree construction for multiscale image representation and processing

With the continuous growth of sensor performances, image analysis and processing algorithms have to cope with larger and larger data volumes. Besides, the informative components of an image might not be the pixels themselves, but rather the objects they belong to. This has led to a wide range of successful multiscale techniques in image analysis and computer vision. Hierarchical representations are thus of first importance, and require efficient algorithms to be computed in order to address real-life applications. Among these hierarchical models, we focus on morphological trees (e.g., min/max-tree, tree of shape, binary partition tree, α-tree) that come with interesting properties and already led to appropriate techniques for image processing and analysis, with a growing interest from the image processing community. More precisely, we build upon two recent algorithms for efficient α-tree computation and introduce several improvements to achieve higher performance. We also discuss the impact of the data structure underlying the tree representation, and provide for the sake of illustration several applications where efficient multiscale image representation leads to fast but accurate techniques, e.g., in remote sensing image analysis or video segmentation.

MedGraph: a graph-based representation and computation to handle large sets of images

In order to process and analyze very large volumes of images, efficient representation and structuring techniques are required. Since, current computing machines can provide large memory size, trading off reasonable amount of memory in order to achieve efficient and parallelizable representation of images is preferable. In this paper, we propose a new structure to represent and store images based on in-memory graph concept. Our method of computation provides a faster execution time than the traditional array-based representation. Each pixel of an image is represented as a one node in the graph. In addition, nodes have pointers to other neighboring nodes (pixels). The structure represents an image as a grid of connected linked lists and each grid is connected to other grids. Using our method, an image can be represented in one of three different representations which are: octal linked list, quadratic linked list, and dual linked list representations. We provide experiments and evaluations using the dual linked list representation as it requires less memory space. We apply our methodology for medical images as a proof of the concept to find a region of interest in an image. We have collected and used real medical images to build and process the graph which we call MedGraph. Our experimental results show that the proposed MedGraph technique improves the searching time for finding a region of interest when compared to the traditional representation. It is worth mentioning here that MedGraph is a generic representation strategy that can be applied to any type of images, although this paper uses medical images as a proof of the concept.

Bag of Features Model Using the New Approaches: A Comprehensive Study

The major challenge in content based image retrieval is the semantic gap. Images are described mainly on the basis of their numerical information, while users are more interested in their semantic content and it is really difficult to find a correspondence between these two sides. The bag of features (BoF) model is an efficient image representation technique for image classification. However, it has some limitations for instance the information loss during the encoding process, an important step of BoF. This is because the encoding is usually done by hard assignment i.e. in vector quantization each feature is encoded by being assigned to a single visual word. Another notorious disadvantage of BoF is that it ignores the spatial relationships among the patches, which are very important in image representation. To address those limitations and enhance the results, novel approaches were proposed at each level of the BoF pipeline. In instance the combination of local and global descriptors for a better description, a soft-assignment encoding manner with a spatial pyramid partitioning for a more informative image representation and a maximum pooling to get the final descriptors. Our work aims to give a detailed version of the BoF, including all the levels of the pipeline, as a support leading to a better comprehension of the approach. We also compare and evaluate the state-of-the-art approaches and find out how these changes at each level of the pipeline could affect the performance and the overall classification results.

Image denoising based on adaptive spatial segmentation and multi-scale correlation in directionlet domain

Directionlet transform DT has become popular over the last few years as an efficient image representation tool due to its fine frequency tiling and directional vanishing moments along any two directions. A novel denoising algorithm based on DT is proposed here for images corrupted with Gaussian noise. The image is first spatially segmented based on the content directionality. Then an undecimated version of DT is applied to effectively capture the directional features and edge information of these segments. The DT coefficients so obtained are then modelled using a bivariate heavy tailed Gaussian distribution and the noise free coefficients are computed using MAP estimator. By employing bivariate probability distribution, the heavy-tail behaviour of natural images is accurately modelled and the interscale properties of DT coefficients are properly exploited. In addition, the local variance parameter of the model is estimated based on classification of DT coefficients within a particular scale using context modelling. Due to this the intrascale dependency of the directionlet coefficients is also well exploited in the enhancement process. The proposed algorithm is competitive with the existing algorithms with better results in terms of output peak signal-to-noise ratio while having comparable computational complexity. It exhibits good capability to preserve edges, contours and textures especially in images with abundant high frequency contents.

Blind Restoration Method for Satellite Images using Memetic Algorithm

Restoration of degraded satellite images are in demand. The sources of degradation can be aliasing, blur, noise and atmospheric turbulence, which are usually an ill-posed in nature. This paper introduces Memetic algorithm for image restoration. Previous restoration techniques have been investigated, but except for certain special cases the maximum cases solve the resulting criterion approximately only. So there is a requirement of more demanding optimization methods. A Memetic algorithm herein proposed give efficient image representation by using hill climbing method for population initialization and using extended neighborhood search. The algorithm is performed on a quickbird test satellite image for the optimization of result. The optimization codes are written in Matlab. The proposed method shows 25% improvement from degraded image. Competitively our new approach performs better than some best existing methods. to demonstrate the effectiveness of the proposed method, comparisons are given from the existing methods. General Terms Image restoration, Hill Climbing, Fitness Function, Genetic Algorithm

Transferring Deep Convolutional Neural Networks for the Scene Classification of High-Resolution Remote Sensing Imagery

Learning efficient image representations is at the core of the scene classification task of remote sensing imagery. The existing methods for solving the scene classification task, based on either feature coding approaches with low-level hand-engineered features or unsupervised feature learning, can only generate mid-level image features with limited representative ability, which essentially prevents them from achieving better performance. Recently, the deep convolutional neural networks (CNNs), which are hierarchical architectures trained on large-scale datasets, have shown astounding performance in object recognition and detection. However, it is still not clear how to use these deep convolutional neural networks for high-resolution remote sensing (HRRS) scene classification. In this paper, we investigate how to transfer features from these successfully pre-trained CNNs for HRRS scene classification. We propose two scenarios for generating image features via extracting CNN features from different layers. In the first scenario, the activation vectors extracted from fully-connected layers are regarded as the final image features; in the second scenario, we extract dense features from the last convolutional layer at multiple scales and then encode the dense features into global image features through commonly used feature coding approaches. Extensive experiments on two public scene classification datasets demonstrate that the image features obtained by the two proposed scenarios, even with a simple linear classifier, can result in remarkable performance and improve the state-of-the-art by a significant margin. The results reveal that the features from pre-trained CNNs generalize well to HRRS datasets and are more expressive than the low- and mid-level features. Moreover, we tentatively combine features extracted from different CNN models for better performance.

Orthogonal Projective Sparse Coding for image representation

We consider the problem of image representation for visual analysis. When representing images as vectors, the feature space is of very high dimensionality, which makes it difficult for applying statistical techniques for visual analysis. One then hope to apply matrix factorization techniques, such as Singular Vector Decomposition (SVD) to learn the low dimensional hidden concept space. Among various matrix factorization techniques, sparse coding receives considerable interests in recent years because its sparse representation leads to an elegant interpretation. However, most of the existing sparse coding algorithms are computational expensive since they compute the basis vectors and the representations iteratively. In this paper, we propose a novel method, called Orthogonal Projective Sparse Coding (OPSC), for efficient and effective image representation and analysis. Integrating the techniques from manifold learning and sparse coding, OPSC provides a sparse representation which can capture the intrinsic geometric structure of the image space. Extensive experimental results on real world applications demonstrate the effectiveness and efficiency of the proposed approach.