Image Feature Detectors and Descriptors

Feature detection, description and matching are essential components of various computer vision applications, thus they have received a considerable attention in the last decades. Several feature detectors and descriptors have been proposed in the literature with a variety of definitions for what kind of points in an image is potentially interesting (i.e., a distinctive attribute). This chapter introduces basic notation and mathematical concepts for detecting and describing image features. Then, it discusses properties of perfect features and gives an overview of various existing detection and description methods. Furthermore, it explains some approaches to feature matching. Finally, the chapter discusses the most used techniques for performance evaluation of detection and description algorithms.

Fast and accurate image feature detectors are an important challenge in computer vision as they are the basis for high-level image processing analysis and understanding. However, image feature detectors cannot be easily applied in real-time embedded computing scenarios, such as autonomous robots and vehicles, mainly due to the fact that they are time consuming and require considerable computational resources. For embedded and low power devices, speed and memory efficiency is of main concern, and therefore, there have been several recent attempts to improve this performance gap through dedicated hardware implementations of feature detectors. Thanks to the fine grain massive parallelism and flexibility of software-like methodologies, reconfigurable hardware devices, such as Field Programmable Gate Arrays (FPGAs), have become a common choice to speed up computations. In this chapter, a review of hardware implementations of feature detectors using FPGAs targeted to embedded computing scenarios is presented. The necessary background and fundamentals to introduce feature detectors and their mapping to FPGA-based hardware implementations are presented. Then we provide an analysis of some relevant state-of-the-art hardware implementations, which represent current research solutions proposed in this field. The review addresses a broad range of techniques, methods, systems and solutions related to algorithm-to-hardware mapping of image interest point detectors. Our goal is not only to analyze, compare and consolidate past research work but also to appreciate their findings and discuss their applicability. Some possible directions for future research are presented.

This chapter addresses many problems: different types of sensors, systems and methods from the literature are briefly revised, in order to give a recipe for designing intelligent vehicle systems based on computer vision. Many computer vision or related problems are addressed, like segmentation, features extraction and selection, fusion and classification. Existing solutions are investigated and three different data-bases are presented to perform typical experiments. Features extraction is aimed for finding pertinent features to encode information about possible obstacles from the road. Feature selection schemes are further used to compact the feature vector in order to decrease the computational time. Finally, several approaches to fuse visible and infrared images are used to increase the accuracy of the monomodal systems.

Image feature extraction is instrumental for most of the best-performing algorithms in computer vision. However, it is also expensive in terms of computational and memory resources for embedded systems due to the need of dealing with individual pixels at the earliest processing levels. In this regard, conventional system architectures do not take advantage of potential exploitation of parallelism and distributed memory from the very beginning of the processing chain. Raw pixel values provided by the front-end image sensor are squeezed into a high-speed interface with the rest of system components. Only then, after deserializing this massive dataflow, parallelism, if any, is exploited. This chapter introduces a rather different approach from an architectural point of view. We present two Application-Specific Integrated Circuits (ASICs) where the 2-D array of photo-sensitive devices featured by regular imagers is combined with distributed memory supporting concurrent processing. Custom circuitry is added per pixel in order to accelerate image feature extraction right at the focal plane. Specifically, the proposed sensing-processing chips aim at the acceleration of two flagships algorithms within the computer vision community: the Viola-Jones face detection algorithm and the Scale Invariant Feature Transform (SIFT). Experimental results prove the feasibility and benefits of this architectural solution.

The rapid increasing of remote sensing (RS) data in many applications ignites a spark of interest in the process of satellite image matching and registration. These data are collected through remote sensors then processed and interpreted by means of image processing algorithms. They are taken from different sensors, viewpoints, or times for many industrial and governmental applications covering agriculture, forestry, urban and regional planning, geology, water resources, and others. In this chapter, a feature-based registration of optical and radar images from same and different sensors using invariant local features is presented. The registration process starts with the feature extraction and matching stages which are considered as key issues when processing remote sensing data from single or multi-sensors. Then, the geometric transformation models are applied followed by the interpolation method in order to get a final registered version. As a pre-processing step, speckle noise removal is performed on radar images in order to reduce the number of false detections. In a similar fashion, optical images are also processed by sharpening and enhancing edges in order to get more accurate detections. Different blob, corner and scale based feature detectors are tested on both optical and radar images. The list of tested detectors includes: SIFT, SURF, FAST, MSER, Harris, GFTT, ORB, BRISK and Star. In this work, five of these detectors compute their own descriptors (SIFT, SURF, ORB, BRISK, and BRIEF), while others use the steps involved in SIFT descriptor to compute the feature vectors describing the detected keypoints. A filtering process is proposed in order to control the number of extracted keypoints from high resolution satellite images for a real time processing. In this step, the keypoints or the ground control points (GCPs) are sorted according to the response strength measured based on their cornerness. A threshold value is chosen to control the extracted keypoints and finalize the extraction phase. Then, the pairwise matches between the input images are calculated by matching the corresponding feature vectors. Once the list of tie points is calculated, a full registration process is followed by applying different geometric transformations to perform the warping phase. Finally and once the transformation model estimation is done, it is followed by blending and compositing the registered version. The results included in this chapter showed a good performance for invariant local feature detectors. For example, SIFT, SURF, Harris, FAST and GFTT achieve better performance on optical images while SIFT gives also better results on radar images which suffer from speckle noise. Furthermore, through measuring the inliers ratios, repeatability, and robustness against noise, variety of comparisons have been done using different local feature detectors and descriptors in addition to evaluating the whole registration process. The tested optical and radar images are from RapidEye, Pléiades, TET-1, ASTER, IKONOS-2, and TerraSAR-X satellite sensors in different spatial resolutions, covering some areas in Australia, Egypt, and Germany.

Video summarization is a task which aims at presenting the contents of a video to the user in a succinct manner so as to reduce the retrieval and browsing time. At the same time sufficient coverage of the contents is to be ensured. A trade-off between conciseness and coverage has to be reached as these properties are conflicting to each other. Various feature descriptors have been developed which can be used for redundancy removal in the spatial and temporal domains. This chapter takes an insight into the various strategies for redundancy removal. A method for intra-shot and inter-shot redundancy removal for static video summarization is also presented. High values of precision and recall illustrate the efficacy of the proposed method on a dataset consisting of videos with varied characteristics.

Dactylology or Finger Spelling is popularly known as sign speech, is a kind of gesture based language used by the deaf and dumb people to communicate with themselves or with other people in and around them. In many ways FingerSpelling provides a connection between the sign and oral language. Dactylology can also be used for secret communication or can be used by the security personnel to communicate secretly with their counterpart. In the proposed work a two phase encryption technique has been proposed wherein the first phase a ‘Gesture Key’, generated from Indian Sign Language in real time has been used for encrypting the Region of Interests (ROIs) and in the second phase a session key has been used to encrypt the partially encrypted image further. The experimental results show that the scheme provides significant security improvement without compromising the image quality. The speed of encryption and decryption process is quite good. The Performance of the proposed scheme is compared with the few other popular encryption methods to establish the relevance of the work.

Spectral imaging has received a great deal of attention recently. Spectral reflectance observed from object surfaces provides crucial information in computer vision and image analysis which include the essential problems of feature detection, image segmentation, and material classification. The estimation of spectral reflectance is affected by several illumination factors such as shading, gloss, and specular highlight. The spectral invariant representations for dielectric materials only, for these factors, are inadequate for other characteristic materials like metal. In this chapter, a spectral invariant representation is introduced for obtaining reliable spectral reflectance images. The invariant formulas for spectral images of natural objects preserve spectral information and are invariant to highlights, shading, surface geometry, and illumination intensity. As an application, a material classification method is presented based on the invariant representation, which results in reliable segmentations for natural scenes and raw circuit boards spectral images.

In image analysis, segmentation is considered one of the most important steps. Segmentation by searching threshold valuesassumes that objects in a digital image can be modeled through distinct gray level distributions. In this chapter it is proposed the use of a bio-inspired algorithm, called Allostatic Optimisation (AO), to solve the multi threshold segmentation problem. Our approach considers that an histogram can be approximated by a mixture of Cauchy functions, whose parameters are evolved by AO. The contributions of this chapter are on three fronts, by using: a Cauchy mixture to model the original histogram of digital images, the Hellinger distance as an objective function, and AO algorithm. In order to illustrate the proficiency and robustness of the proposed approach, it has been compared to the well-known Otsu method, over several standard benchmark images.

With the use of computers and Internet in every major activity of our society, security is increasingly important. Biometric recognition is not only challenging but also computationally demanding. This chapter aims develop an iris biometric system. The iris has the advantages of uniqueness, stableness, anti-spoof, non-invasiveness and efficiency and could be applied in almost every area (banking, forensics, access control, etc.). The performance of a biometric classification system is largely depending on the techniques used for feature extraction. Inspired by the biological plausibility of ordinal measures, we propose their employment for iris representation and recognition. Qualitative measurement, associated to the relative ordering of different characteristics, is defined as ordinal measurement. Besides the proposing of a novel, fast and robust, ordinal based feature extraction method, the chapter also considers the problem of designing the decision making model so as to obtain an efficient and effective biometric system. In the literature, there are different approaches for iris recognition, nevertheless, there are still challenging open problems in improving the accuracy, robustness, security and ergonomics of biometric systems.

Detection of the center of a fovea and its boundary from a retinal image is a challenging task due to the irregularity of the avascular foveal region. Several attempts have been made before in order to detect the foveal region and its boundary from fluorescein angiographic images. The irregularity and large variation in the human retinal images made the task increasingly difficult. In this current work funds images were considered instead of fluorescein angiographic images in order to ensure the applicability of the proposed algorithm in both biomedical and biometric analysis. A robust fuzzy-rule based image segmentation algorithm has been developed in order to extract the foveal region from a wide variety of images from different persons. Detection of foveal region comprised of locating the geometric center and extracting the boundary. The geometric center was evaluated by weighted averaging the grey scale intensities obtained from implementing the current algorithm. This was followed by applying gradient vector flow (GVF) based active contour technique in order to extract the boundary of the foveal region. The algorithm was applied on a several retinal images acquired from different persons with a very good success rate. The present work is considered to be an important contribution in intelligent image analysis of human retina since it incorporates a robust “fuzzy-rule” in extracting foveal region. Similar approach has not been adopted in the literature. The proposed algorithm is seen to be versatile in analyzing a wide range of retinal images.

Registration of digital elevation models is a vital step in fusing sensor data. In this chapter, we present a robust topological framework for entropic image registration using Morse singularities. The core idea behind our proposed approach is to encode a digital elevation model into a set of nondegenerate singular points, which are the maxima, minima and saddle points of the Morse height function. An information-theoretic dissimilarity measure between the Morse features of two misaligned digital elevation models is then maximized to bring the elevation data into alignment. In addition, we show that maximizing this dissimilarity measure leads to minimizing the total length of the joint minimal spanning tree of two misaligned digital elevation data models. Illustrating experimental results are presented to show the efficiency and registration accuracy of the proposed framework compared to existing entropic approaches.

The present work focuses on the aspect of textural variations exhibited by primary benign and primary malignant focal liver lesions. For capturing these textural variations of benign and malignant liver lesions, texture features are computed using statistical methods, signal processing based methods and transform domain methods. As an application of texture description in medical domain, an efficient CAD system for primary benign i.e., hemangioma (HEM) and primary malignant i.e., hepatocellular carcinoma (HCC) liver lesions based on texture features derived from B-Mode liver ultrasound images of Focal liver lesions has been proposed in the present study. The texture features have been computed from the inside regions of interest (IROIs) i.e., from the regions inside the lesion and one surrounding region of interest (SROI) for each lesion. Texture descriptors are computed from IROIs and SROIs using six feature extraction methods namely, FOS, GLCM, GLRLM, FPS, Gabor and Laws’ features. Three texture feature vectors (TFVs) i.e., TFV1 consists of texture features computed from IROIs, TFV2 consists of texture ratio features (i.e., texture feature value computed from IROI divided by texture feature value computed from corresponding SROI) and TFV3 computed by combining TFV1 and TFV2 (IROIs texture features + texture ratio features) are subjected to classification by SVM and SSVM classifiers. It is observed that the performance of SSVM based CAD system is better than SVM based CAD system with respect to (a) overall classification accuracy (b) individual class accuracy for atypical HEM class and (c) computational efficiency. The promising results obtained from the proposed SSVM based CAD system design indicates its usefulness to assist radiologists for differential diagnosis between primary benign and primary malignant liver lesions.

It has been strongly advocated that increase in density of breast tissue is strongly correlated with the risk of developing breast cancer. Accordingly change in breast tissue density pattern is taken seriously by radiologists. In typical cases, the breast tissue density patterns can be easily classified into fatty, fatty-glandular and dense glandular classes, but the differential diagnosis between atypical breast tissue density patterns from mammographic images is a daunting challenge even for the experienced radiologists due to overlap of the appearances of the density patterns. Therefore a CAD system for the classification of the different breast tissue density patterns from mammographic images is highly desirable. Accordingly in the present work, exhaustive experiments have been carried out to evaluate the performance of statistical features using PCA-kNN, PCA-PNN, PCA-SVM and PCA-SSVM based CAD system designs for two-class and three-class breast tissue density classification using mammographic images. It is observed that for two-class breast tissue density classification, the highest classification accuracy of 94.4 % is achieved using only the first 10 principal components (PCs) derived from statistical features with the SSVM classifier. For three-class breast tissue density classification, the highest classification accuracy of 86.3 % is achieved using only the first 4 PCs with SVM classifier.