Explore chapters and articles related to this topic
Vehicle detection using faster R-CNN
Published in Brij B. Gupta, Nadia Nedjah, Safety, Security, and Reliability of Robotic Systems, 2020
Kaustubh V. Sakhare, Pallavi B. Mote, Vibha Vyas
Various approaches are involved in the vehicle detection methods, which include motion, hand-crafted feature, and CNN-based approaches. Success in classification and detection of the vehicle using hand-crafted feature-based approach depend upon the skill test of the programmer and cannot meet the optimum feature representation [7]. In the context of vehicle detection, hand-crafted features are extracted for the learning model. The object detection and localization methods are analyzed based on the quality of the feature extraction. The shape is one of the prominent features used to represent the object. Few driving factors for selection of the shape descriptors are it should be invariant to the translation, rotation, and scaling. Shape descriptors should be invested insightfully in object detection [2,15]. A large number of techniques have been proposed for describing shapes in object detection wherein the points of interest are discerned out of the images and compared to those with the ones registered from dataset images to find the object of interest [2]. This part of interest inside the image is normally treated as the feature. Compared to other detectors, speeded up robust features (SURF) and SIFT feature detectors are robust. Even a better localization quality as expected than others while considering real-time vehicle detection. It quickly detects and classifies the objects as these are rotation and scale-invariant. These methods are employed in feature extraction for a long time.
Advanced Sensors and Vision Systems
Published in Marina Indri, Roberto Oboe, Mechatronics and Robotics, 2020
Many kinds of methods for extraction and matching of feature descriptors have been proposed: Harris corner [28], distinctive image features from SIFT [29], speeded up robust features (SURF) [30], gradient location and orientation histogram (GLOH) [31], histograms of oriented gradients (HOG) [32], etc. SIFT was first presented by David G Lowe, and it has been successfully applied in the pattern recognition and matching field. Practical experiments have shown that the SIFT algorithm is very invariant and robust in scaling, rotation, and affine transformation for feature detection and matching, although it is more expensive in computational processing. According to these conclusions and experimental results, the SIFT feature descriptor is utilized to detect and match for correspondence points to estimate motion and localization. The SIFT algorithm processing is briefly described through several steps as follows: scale space extreme detection, accurate key point localization, orientation assignment, and key point descriptor [33, 34].
Advances in Unconstrained Handprint Biometrics
Published in Karm Veer Arya, Robin Singh Bhadoria, The Biometric Computing, 2019
Gaurav Jaswal, Amit Kaul, Ravinder Nath
UR-SIFT: To perform image matching by using the set of local interest points is an important aspect of local texture descriptor-based approaches. Local image features like blobs, points and micro-image regions are unique and facilitate such schemes to better handle varying illumination, translation and scale (Jaswal et al. 2017a). The SIFT algorithm and its variants have been effectively applied in different computer vision and image processing applications. However, standard SIFT causes problems with the multi-source remote sensing images, particularly unable to extract required number of feature points because such images composed information over a large range of variation on frequencies. In this paper, a strongest variant of SIFT known as uniform robust SIFT (UR-SIFT) is employed for image matching (Sedaghat et al. 2011). The UR-SIFT is based on an idea of selecting high-quality SIFT features with a uniform distribution in both the scale and image spaces. It has been observed that UR-SIFT extracts more key points than standard SIFT and perform matching efficiently, as shown in Figure 2.12.
Recognition of Indian Sign Language Using ORB with Bag of Visual Words by Kinect Sensor
Published in IETE Journal of Research, 2022
Jayesh Gangrade, Jyoti Bharti, Anchit Mulye
This method, Scale Invariant Feature Transform (SIFT) provides features of the image, which are scale, rotation, and illumination invariant [38]. The various hand postures in sign language are oriented which eradicate the chance of rotation in any image. All the images in proposed method are taken using the Kinect sensor, and therefore images do not vary in the scale. The resulted key points by SIFT are very less and also very divergent for the same type of images. The relative distance between any two key points differs much radically for different images of the same type. SIFT algorithm usually very efficient for low-resolution images, and its features are invariant to scale and rotation. It works in four stages. In the first step, the Gaussian filter applied to measure the location of potential interest point in the image. Then, the low contrast point is ignored. Each key point is assigned by local image features. The size of the feature vector determined based on several histograms and the number of bins in each histogram. In the original implementation of SIFT, include a 4-by-4 patch of the histogram with eight bins. It generates the 128-dimensional feature vector. Figure 8 shows various key points detected for Sign 1, 2, 5, and 8. The Table 3 shows the average number of features generated by SIFT for 0–9 ISL signs.
Research status and prospect of visual image feature point detection in body measurements
Published in The Journal of The Textile Institute, 2022
Wenqian Feng, Yanli Hu, Xinrong Li, Yuzhuo Li
The SIFT algorithm was proposed by David Lowe in 1999, and then refined in 2004 (Lowe, 2004). SIFT is a widely used feature point recognition methods, and has been successfully applied to computer vision algorithms such as target detection, target tracking, and large-scale image retrieval (Acharya et al., 2018). The SIFT algorithm is divided into four steps: scale space extreme value detection, feature point location, determination of feature point direction, and construction of feature point descriptors. For the detection of the feature points of a 2D image of body measurements, it is only necessary to focus on the first two steps of the algorithm and locate the feature points. Using the Gaussian blur operation in different scale spaces, that is, a Gaussian kernel and the image convolution operation, the image interval points are sampled and an image pyramid is constructed. This Gaussian pyramid is then used to subtract the upper and lower layers in each group to obtain the Gaussian difference image, and then the difference-of-Gaussians (DOG) function is applied to detect and locate key points (Xu et al., 2019). The advantage of this algorithm is that it is not affected by the light level. In the case of complex, noisy scenes, it detects prominent feature points such as edge points, corner points, and points with different brightness values in bright and dark areas.
Content-based image retrieval: A review of recent trends
Published in Cogent Engineering, 2021
Ibtihaal M. Hameed, Sadiq H. Abdulhussain, Basheera M. Mahmmod
Scale-invariant feature transform (SIFT) is one of the most widely used local descriptors introduced by David Lowe (Low, 2004), which contains a detector and a descriptor for key points. SIFT is robust against image rotation and image scaling, but it performs poorly in matching at high dimensions and need a fixed-size vector for encoding to perform image similarity checking. In image retrieval, SIFT has two drawbacks: it uses a large amount of memory and has a high computational cost (Montazer & Giveki, 2015). Soltanshahi et al. proposed (Montazer & Giveki, 2015) a method for CBIR based on the use of SIFT and local derivative pattern (LDP) to construct the feature descriptor. To overcome its limitation of high memory usage and computation cost, the authors proposed two methods to reduce SIFT dimensionality. The proposed system was tested using four datasets and proved its high retrieval performance for images that contain an object. However, it needs to be improved for nature images.