Wyniki wyszukiwania - BazTech

1

Overcoming Overfitting Challenges with HOG Feature Extraction and XGBoost-Based Classification for Concrete Crack Monitoring

Barkiah Ida, Sari Yuslena

International Journal of Electronics and Telecommunications

|

2023

|

Vol. 69, No. 3

571--577

EN

This study proposes a method that combines Histogram of Oriented Gradients (HOG) feature extraction and Extreme Gradient Boosting (XGBoost) classification to resolve the challenges of concrete crack monitoring. The purpose of the study is to address the common issue of overfitting in machine learning models. The research uses a dataset of 40,000 images of concrete cracks and HOG feature extraction to identify relevant patterns. Classification is performed using the ensemble method XGBoost, with a focus on optimizing its hyperparameters. This study evaluates the efficacy of XGBoost in comparison to other ensemble methods, such as Random Forest and AdaBoost. XGBoost outperforms the other algorithms in terms of accuracy, precision, recall, and F1-score, as demonstrated by the results. The proposed method obtains an accuracy of 96.95% with optimized hyperparameters, a recall of 96.10%, a precision of 97.90%, and an F1-score of 97%. By optimizing the number of trees hyperparameter, 1200 trees yield the greatest performance. The results demonstrate the efficacy of HOG-based feature extraction and XGBoost for accurate and dependable classification of concrete fractures, overcoming the overfitting issues that are typically encountered in such tasks.

2

Integration of feature extraction, attribute combination and image segmentation for object delineation on seismic images

Khayer Keyvan, Hosseini Fard Esmail, Roshandel Kahoo Amin, Soleimani Monfared Mehrdad, Ahmadyfard Alireza

Acta Geophysica

|

2023

|

Vol. 71, no. 1

275--292

EN

Automatic geological interpretation, specifically modeling salt dome and fault detection, is controversial task on seismic images from complex geological media. In advanced techniques of seismic interpretation and modeling, various strategies are utilized for combination and integration different information layers to obtain an image adequate for automatic extraction of the object from seismic data. Efficiency of the selected feature extraction, data integration and image segmentation methods are the most important parameters that affect accuracy of the final model. Moreover, quality of the seismic data also affects confidence of the selected seismic attributes for integration. The present study proposed a new strategy for efficient delineation and modeling of geological objects on the seismic image. The proposed method consists of extraction specific features by the histogram of oriented gradients (HOG) method, statistical analysis of the HOG features, integration of features through hybrid attribute analysis and image classification or segmentation. The final result is a binary model of the target under investigation. The HOG method here modified accordingly for extraction of the related features for delineation of salt dome and fault zones from seismic data. The extracted HOG parameter then is statically analyzed to define the best state of information integration. The integrated image, which is the hybrid attribute, then is used for image classification, or image segmentation by the image segmentation method. The seismic image labeling procedure performs on the related seismic attributes, evaluated by the extracted HOG feature. Number of HOG feature and the analyzing parameters are also accordingly optimized. The final image classification then is performed on an image which contains all the embedded information on all the related textural conventional and statistical attributes and features. The proposed methods here apply on four seis mic data examples, synthetic model of salt dome and faults and two real data that contain salt dome and fault. Results have shown that the proposed method can more accurately model the targets under investigation, compared to advanced extracted attributes and manual interpretations.

3

Analysis of Clothing Image Classification Models: A Comparison Study between Traditional Machine Learning and Deep Learning Models

Xu Jun, Wei Yumeng, Wang Aichun, Zhao Heng, Lefloch Damien

Fibres & Textiles in Eastern Europe

|

2022

|

Vol. 30, no. 5

66--78

EN

Clothing image in the e-commerce industry plays an important role in providing customers with information. This paper divides clothing images into two groups: pure clothing images and dressed clothing images. Targeting small and medium-sized clothing companies or merchants, it compares traditional machine learning and deep learning models to determine suitable models for each group. For pure clothing images, the HOG+SVM algorithm with the Gaussian kernel function obtains the highest classification accuracy of 91.32% as compared to the Small VGG network. For dressed clothing images, the CNN model obtains a higher accuracy than the HOG+SVM algorithm, with the highest accuracy rate of 69.78% for the Small VGG network. Therefore, for end-users with only ordinary computing processors, it is recommended to apply the traditional machine learning algorithm HOG+SVM to classify pure clothing images. The classification of dressed clothing images is performed using a more efficient and less computationally intensive lightweight model, such as the Small VGG network.

4

Extracting the symmetry of the human face from digital photographs

Renero-C F. J., Romero-H R. A., Peregrina-B H.

Bio-Algorithms and Med-Systems

|

2017

|

Vol. 13, no. 2

103--109

EN

By defining a midline and selecting six pairs of the landmarks of the human face on digital photographs, we extracted the symmetry of the human face by means of digital techniques. As a first approach to the symmetry of the human face, the distances and the tilts from the midline, between similar landmarks, were computed and averaged, respectively. The procrustes analysis and the histogram of oriented gradients (HOG), applied on patches on the six pairs of the landmarks of the human face, were used as a second approach to the symmetry of the human face. To have a better estimation of the symmetry of the whole human face, the photographs in grayscale and color were cut on pairs of strips, equally spaced from the midline, and then the strips were compared by the HOG feature extractor. The symmetry of the human face was extracted from 89 photographs of human faces (37 females and 52 males, ages 28.67 } 6.65 and 35.65 } 12.2 years, respectively). The HOG feature extractor applied on strips for the photographs in color and grayscale provided more confident values for the symmetry of the human face, which was well correlated with the assigned value by the photographers and physiotherapists. Also, an experiment was performed to evaluate the attractiveness as a function of the human face symmetry; thus, two groups of men and women were asked to sort digital photographs of women and men according to the attractiveness of women/men on the photographs. The results show that the most selected digital photographs were those with the highest symmetry scores.

5

FPGA Implementation of Multi-scale Face Detection Using HOG Features and SVM Classifier

Drożdż M., Kryjak T.

Image Processing & Communications

|

2016

|

Vol. 21, no. 3

27--44

EN

In this paper an FPGA based embedded vision system for face detection is presented. The sliding detection window, HOG+SVM algorithm and multi-scale image processing were used and extensively described. The applied computation parallelizations allowed to obtain real-time processing of a 1280 × 720 @ 50Hz video stream. The presented module has been verified on the Zybo development board with Zynq SoC device from Xilinx. It can be used in a vast number of vision systems, including diver fatigue monitoring.

6

Human Detection in Thermal Images Using Low-level Features

Budzan S.

Measurement Automation Monitoring

|

2015

|

Vol. 61, No. 6

191--194

EN

In this work the human detection method in infrared has been presented. The proposed solution focuses on the use low-level features and detecting parts of the human body. Low–level processing is based on modified HOG (Histogram of Oriented Gradients) algorithm. First, the only squared cells have been used, also calculation of the gradient has been improved. Next, the model of the head from the dataset IR (Infra Red) images has been created, also the model of the human body. Finally, the probability matrix has been examined using minimal distance classifier. The novelty of the proposed solution focuses on the combination of the pixel-gradient and body parts processing, also three stage classification process (head modelling, human modelling and classifier), which has been proposed to reduce the false detection. The experiments were performed on self-created IR images database, which contains images with most of the possible difficult situations such as overlapped people, different pose, small and high resolution of the people. The performance of the proposed algorithm was evaluated using Precision and Recall quality measure.

7

Research of Image Features for Classification of Wear Debris

Juránek R., Machalik S., Zemčik P.

Machine Graphics and Vision

|

2011

|

Vol. 20, No. 4

479-493

EN

The wear debris of engineering equipment (such as combustion engines, gearboxes, etc.) consists of metal particles which can be obtained from lubricants used in the equipment. The analysis of wear particles is very important for early detection and prevention of failures. The analysis is often done using classication of individual wear particles obtained by analytical ferrography. In this paper, we present a study of feature extraction methods for a classication of wear particles based on visual similarity. The main contribution of the paper is the comparison of nine selected feature types in the context of three state-of-the-art learning models. Another contribution is the large public database of particle images which can be used for further experiments. The paper describes the dataset, presents the methods of classication, demonstrates the experimental results, and draws conclusions.