Histogram of Oriented Gradients (HOG)
Histogram of Oriented Gradients (HOG) is a feature extraction method used for object detection and recognition in the fields of computer vision and image processing. The principle of HOG is to capture the edge and gradient direction information in an image and represent the object features based on this information. It is mainly used in the following applications: 1.
Object Detection: HOG is used to detect objects in images, for example, people detection or car detection, etc. HOG features are useful for detection because they capture information about the shape and edges of objects. 2.
Object Recognition: HOG features are used in object recognition tasks because they help to represent the appearance information of objects. These include, for example, handwriting recognition, face recognition, and animal recognition.
The basic idea of HOG consists of the following steps
1. image gradient calculation: For each pixel in the image, the gradient (gradient) is calculated. Usually, a Sobel or Scharr filter is used to compute the gradient components in the x- and y-directions for each pixel, which provides the edge information in the image.
2. cell-by-cell gradient histogram computation: The image is divided into small blocks called cells, and within each cell the gradient direction is represented as a histogram. The gradient direction of a pixel in a cell is assigned to a histogram bin, and thus the edge direction information is aggregated for each cell. 3.
3. Block-by-Block Normalization: Combines the histogram information of cells into regions called blocks. The information between neighboring cells is combined, the features within the block are computed, and normalization is performed on a block-by-block basis, resulting in features that are robust to changes in light.
4. feature vector formation: The information in the normalized blocks is concatenated to form the final feature vector. This feature vector is used to represent the object as a HOG feature.
HOG is used especially in combination with machine learning algorithms such as SVM (Support Vector Machine) to achieve high performance in object detection and object recognition tasks. Its relatively low computational cost also makes it suitable for real-time applications. However, it is not robust against rotation or viewpoint changes, so care should be taken especially when rotational invariance is required.
Specific procedures for Histogram of Oriented Gradients (HOG)
This section describes the specific procedures of Histogram of Oriented Gradients (HOG), a feature extraction method for object detection and recognition, and the calculation procedure is as follows
1. image preprocessing:
Before image processing, the input image is resized to the required size and, if necessary, grayscaled (monochromatized); HOG is usually applied to monochrome images.
2. gradient computation:
For each pixel in the image, the gradient component (the rate of change of luminance in the horizontal and vertical directions) is calculated in the x- and y-directions. A common method is to use a Sobel or Scharr filter. This provides edge information in the image.
3. gradient intensity and angle computation:
From the x- and y-direction components of the gradient, the gradient intensity (magnitude of the gradient vector) and gradient angle (direction of the gradient vector) at each pixel are calculated. The gradient angle is usually expressed in the range of 0 to 180 degrees.
4. cell definition:
The image is divided into small square blocks called cells. The pixels within a cell are later used to compute a histogram. Typical cell sizes are 16×16 pixels or 8×8 pixels, etc.
5. Gradient histogram calculation per cell:
For each pixel in each cell, a histogram is computed using the gradient intensity and gradient angle. The histogram is divided into a series of angle bins and the gradient intensities within the angle range are added to each bin.
6. block definition:
Cells are combined to form a large square area called a block. A block is a set of cells, and the cell information within the block is later used for normalization. Blocks are usually arranged so that adjacent cells overlap.
7. Normalization per block:
Cell information within a block is normalized. It is usually normalized in the L2 norm (Euclidean norm), which results in features that are robust to light variations.
8. feature vector formation:
The information in the normalized blocks is concatenated block by block to form the final feature vector. This feature vector is used to represent the object as HOG features.
9. application to machine learning models:
The HOG feature vectors are typically input to a machine learning model, such as a Support Vector Machine (SVM), which is used for object detection or object recognition tasks. The model analyzes the feature vector to detect or recognize the presence of an object.
This technique is useful for many object detection and recognition tasks because it emphasizes object shape and edge information and provides features that are robust to the background.
For an example implementation of Histogram of Oriented Gradients (HOG)
To implement Histogram of Oriented Gradients (HOG), the following steps can be followed using the Python and OpenCV libraries. The following example shows how to use OpenCV’s HOG detector to compute HOG features and perform object detection.
Install OpenCV: Install OpenCV to compute HOG features.
pip install opencv-pythonCreate HOG Detector: Create an OpenCV HOG detector to compute HOG features.
import cv2
# Initialize HOG Detector
hog = cv2.HOGDescriptor()Image loading: Load images for calculating HOG features.
image = cv2.imread('sample.jpg')Calculation of HOG features: Calculate HOG features using the HOG detector.
# HOG特徴を計算
features = hog.compute(image)HOG features are stored in features.
Object detection: To detect specific objects, object detection is performed using HOG features. The following is an example of object (e.g., car) detection.
# SVM model for HOG detector
hog.setSVMDetector(cv2.HOGDescriptor_getDefaultPeopleDetector())
# Perform object detection
detected_objects, _ = hog.detectMultiScale(image)detected_objects contains the location information of the detected objects.
Draw the result: A rectangle is drawn on the original image using the location information of the detected objects.
for (x, y, w, h) in detected_objects:
cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
# View resulting images
cv2.imshow('Object Detection', image)
cv2.waitKey(0)
cv2.destroyAllWindows()This completes the basic implementation of object detection using HOG features, which can be used in a variety of object detection tasks and are widely used especially in tasks such as row person detection. It is important to adjust HOG detector settings and SVM model selection according to the object and task of object detection.
Cahllemge of Histogram of Oriented Gradients (HOG)
Histogram of Oriented Gradients (HOG) is a useful feature extraction method, but there are several challenges and limitations. The main challenges of HOG are listed below.
1. lack of robustness against rotation:
HOG features are not robust against object rotation. When objects are viewed at different angles, HOG features do not match, resulting in poor object detection performance.
2. coping with scale changes:
When the size of an object changes, HOG features may have difficulty coping. They are particularly unsuitable for detecting small objects and need to be computed at different scales to detect objects of different scales.
3. sensitivity to shadows and illumination changes:.
HOG features are sensitive to illumination changes and shadows, and changes in illumination conditions in an image can cause fluctuations in HOG features and reduce object detection reliability.
4. effect on background information:
The response of HOG features to background can affect object detection. Especially in the case of complex backgrounds, false positives are more likely to occur.
5. computational cost:
Computing HOG features is computationally expensive. Especially for high-resolution images, the computation time increases and may not be suitable for real-time object detection.
6. generalization to different object classes:
HOG features tend to extract features specific to a particular object class and may be difficult to generalize to other object classes. Detection of different object classes requires different settings and training data.
To overcome these challenges, research has been conducted to improve HOG features or combine them with other feature extraction methods. In particular, development of methods to provide rotation-invariant and scale-invariant features is underway, and deep learning-based approaches (e.g., CNN described in “Overview of CNN and examples of algorithms and implementations”) have been successful in addressing some of the challenges and are sometimes used in combination with HOG.
Strategies to address Histogram of Oriented Gradients (HOG) issues
Measures and improvements to address the challenges of Histogram of Oriented Gradients (HOG) are described below.
1. improving rotation invariance:
One way to improve the rotation invariance of HOG is to compute multiple HOG feature vectors for multi-angle views of an object and combine them. Another possibility is to use scale pyramids to extract HOG features corresponding to multiple scales and rotations.
2. scale invariance enhancement:
One way to deal with scale variation is to compute multi-scale HOG features. One may process images at different resolutions, calculate HOG features for each resolution, and devise a normalization method for the HOG features to obtain features that are robust to scale.
3. coping with illumination changes:
One way to deal with illumination changes is to perform contrast normalization and extraction of illumination-invariant features. This will result in HOG features that are robust to changes in lighting conditions in the image.
4. background information reduction:
To reduce the influence of background information, background subtraction may be applied before object detection, or post-processing methods may be used to reduce false positives in object regions.
5. speed-up:
To reduce computational cost, there are ways to accelerate HOG computations. Examples include using GPUs, parallelization of histogram computation, and partial computation of features.
6. introduction of deep learning:
Deep learning-based object detection models (e.g., CNN-based models) may be easier to address many challenges than HOG. Using deep learning models can achieve a high degree of rotational invariance, scale invariance, and illumination invariance.
7. data augmentation:
Data augmentation techniques can be used to increase training data and improve the generalization performance of HOG-based models.
Reference Information and Reference Books
For details on image information processing, see “Image Information Processing Techniques.
Reference book is “
“
“
“
コメント