Feature Detection

Computer vision relies on image features describing points, edges, objects or colour. The book concerns solely so-called hand-made features contrary to learned features which exist in deep learning methods. Image features can be generally divided into glo

  • PDF / 1,415,680 Bytes
  • 26 Pages / 439.37 x 666.142 pts Page_size
  • 53 Downloads / 235 Views

DOWNLOAD

REPORT


Feature Detection

Computer vision relies on image features describing points, edges, objects or colour. The book concerns solely so-called hand-made features contrary to learned features which exist in deep learning methods. Image features can be generally divided into global and local methods. Global methods extract features from the entire image without dividing into more and less significant areas. To this group, we can include histogram-based algorithms such as histogram of oriented gradients (HOG) or colour coherence vector (CCV) [12, 47]. In most cases, they generate a constant amount of description data which is easier to compare and store, on the other hand, image comparison by histogram-based algorithms gives only a vague similarity for a user. Local feature-based methods try at first to find significant characteristic areas of an image based on Laplacian of Gaussian (LoG) or Difference of Gaussian (DoG) algorithms [25, 64]. And then they generate a description of their neighbourhood. These methods are more accurate, on the other hand, can generate a lot of description data and that amount varies per image. Local feature methods based on keypoints are efficient in similarity detection between images but less in content recognition. Commonly used methods of this kind are SIFT, SURF, ORB, BRIEF, FAST [4, 6, 46, 48, 49].

2.1 Local Features 2.1.1 Scale-Invariant Feature Transform (SIFT) SIFT (Scale-Invariant Feature Transform) is an algorithm used to detect and describe local features of an image. It was presented for the first time in [37] and is now © Springer Nature Switzerland AG 2020 R. Scherer, Computer Vision Methods for Fast Image Classification and Retrieval, Studies in Computational Intelligence 821, https://doi.org/10.1007/978-3-030-12195-2_2

7

8

2 Feature Detection

patented by the University of British Columbia. For each keypoint, which describes the local image feature, we generate a feature vector, that can be used for further processing. The algorithm is immune to changing scale, rotation and light change. SIFT consists of the four main steps [36] 1. Scale-space extreme detection—Extraction of potential keypoints by scanning the entire image, Constructing scale-space, Laplacian approximation by Gaussian blur, 2. Keypoint localization—Selection of stable keypoints (resistant to change of scale and rotation), Removing not important keypoints (noise). 3. Orientation assignment—Finding keypoint orientation resistant to the image transformation, 4. Keypoint descriptor—Generating vectors describing keypoints. During the process of creating scale-space, the image is rescaled (creating octaves) in order to detect the most important and resistant features. After this step, a scale-space pyramid is obtained. This pyramid consists of octaves, sorted from the largest to the smallest octave. In the next stage, Gaussian blur is applied. This step is performed by the following Gaussian operator [37] L(x, y, σ) = G(x, y, σ) ∗ I (x, y),

(2.1)

where L is the output image, G represents the Gaussian operat