Image Processing Techniques and Deep Learning Algorithm for Crack Detection and Classification

One of the initial signs of degradation of a concrete surface is cracks. Cracks may develop in the walls of the building due to many reasons such as seasonal changes and poor quality of materials. In this research both image processing techniques as well as deep learning algorithm have been combined for crack detection and classification. Deep learning algorithm has been adopted in the proposed methodology because of the accurate results and less time taken when compared with other algorithms such as support vector machine and k nearest neighbor algorithm. The computational complexity is also less in the proposed methodology. Two of the preprocessing methods such as filtering and edge detection have been compared and the best methods in terms of Peak signal to Noise ratio and accuracy have been implemented in the proposed methodology. The proposed methodology has been chosen after the comparison of filtering and edge detection methods and those efficient methods have been merged with deep learning algorithm because of its accuracy.

Introduction

The process of detecting cracks in the building walls and also in concrete surfaces is known as crack detection. Destructive testing and non-destructive testing are the two ways to perform crack detection. After the cracks have been detected the dimensions of the cracks should be measured. Human inspection has got many drawbacks such as time consuming and it will be slower than automatic crack detection methods. Accuracy is the main reason behind the adoption of image processing techniques and deep learning algorithm in crack detection. Cracks in building walls or concrete surfaces have been analyzed as an important reference factor of safety evaluation. Thus crack detection in concrete surface plays a very vital role in the maintenance of concrete structures. By using automatic crack detection technology we will be able to overcome the subjectivity of the traditional manual methods.

Crack detection has got some steps associated with it. They are preprocessing, detection and classification. Literature states that smoothing and filtering methods were used for pre-processing and the phase of detection has been carried out by many methods such as Otsu method, statistical approach, threshold method etc. In our proposed methodology classification has been done using deep learning algorithm called convolutional deep learning algorithm because of accurate results.

Image acquisition is a crucial method in visual inspection. Different types of techniques are employed for capturing high intention facsimile. The high resolution pictures may be procured using high resolution scanning. A high resolution camera of great spatial exactitude and color fedility could be preferable for this type of image acquisition, where the sample capturing should be done on altered resolutions and their consistent greyscale image.

This is the primary step of any image processing technique, during this stage, we tend to create the input image compatible for process. There are innumerable distractions that affl ict the input image like illumination variations, noise, backgrounds, variations in image sizes etc. So the fi rst processing phase in external fault detection is noise removal, input image enhancement, which is completed in the preprocessing phase. Also the process of external fault recognition take grey scale image instead of color image, therefore a conversion from color to greyscale is needed for the process, which is done in the preprocessing stage. Images are systemized for further processing in the preprocessing stage

The procedure of segregating an image into its essential areas or objects is called image segmentation. The inputs of segmentation are image but unlike other image processing methods, the outputs of segmentation are attributes i.e. natural qualitative extracted from those images. Also the segmentation process should be stopped as soon as the objects or regions of interest are detected. Different types of segmentation techniques are used for external fault recognition like thresholding, template matching, boundary detection texture matching and so on.

Feature extraction involves reducing the quantity of resources needed to describe large set of information, it can be transferred in to a reduced set of features. Classification algorithms aim at finding similarities in patterns of empirical information. The classification process is based on the features extracted, it classifies the features and makes result. The most commonly used classifiers are neural network classifier, SVM, Bayesian etc.

Deep learning algorithms have been investigated for solving many challenging problems in image processing and classification. In our proposed methodology cracks have been detected and classification has been done using image processing methods such as filtering and canny edge detection. Those results have been integrated with the deep learning algorithm known as convolutional deep learning algorithm. Also a comparison has been done on two of the filtering methods , they were average filter and median filter. Edge detection techniques such as canny edge detection and sobel edge detection were also compared in terms of accuracy and time taken.

In order to ensure the safety and durability of a concrete structure the crack assessment has to be done on a regular basis. Many researchers have studied the automated concrete crack detection method. Some researchers like Abdel-Qader et al. (2003) have acquired the data from structures by using CCTV, laser scanner etc. Several methods have been proposed by the researchers, some of them were, Brilakis et al. (2011) have proposed image processing using edge detection techniques. Deason, J. P et al. (1998) have proposed histogram matching, image filtering and change detection methods. Another method called automatic thresholding valley-emphasis method―a revised version of the Otsu method for detecting small to large defects was introduced by Goedert et al. (2005). Sinha et al. (2006) have analysed classification through neuro-fuzzy network and Khanfar et al. (2003) have proposed the concrete structure defects through fuzzy logic techniques. In addition to their method, neural network and genetic algorithm have been used. In many papers, the inspection methods vary widely from data acquisition to classification and these shows that many algorithms could be used for detecting surface defects.

Researchers have found that it is difficult to apply the inspection algorithm to the structures that are exposed to various weather conditions. If the inspection algorithm is influenced by external conditions, a system engineer should participate in the inspection of structures, because parameter tuning requires expert’s knowledge.

Deep Learning Algorithms for Classification

There are many deep learning algorithms that can be used for classification. Some of the deep learning algorithms are back propagation, fuzzy logic controlled deep neural network algorithm, Fuzzy neural network training algorithm and convolutional neural network algorithm.

Machine learning algorithms have been divided in two, they are supervised learning and unsupervised learning. Supervised learning algorithms are being further divided into classification and regression and unsupervised has been further divided as clustering algorithms. Classification algorithms have been divided into neural networks algorithm, nearest neighbor algorithm, Support Vector Machine algorithm etc.

The main objective of Back Propagation method is adapting synaptic weights in order to minimize an error function. The approach most commonly used for the minimization of the error function is based on the gradient method. Leo et al. (2017) has recommended that fine tuning is a strategy that is commonly found in deep learning. It can be used to greatly improve the performance of stacked auto-encoders. As the back propagation algorithm which is based on descent gradient technique can be extended to apply for an arbitrary number of layers, back propagation algorithm can be used on stacked auto- encoders of arbitrary depth. In their work, to adopt the connections weights were adopted in order to obtain minimal difference between the network output and the desired output. The algorithm is very simple and the output of neural network is evaluated against desired output. Connection between layers will be modified and the process is repeated again until error is small enough if the results are not satisfactory.

Leoet al. (2017) has proposed a fuzzy logic management technique which may be helpful in representing human information in a very specific domain of application and in reasoning there with information to create helpful inferences or actions. A symbolic logic system consists of 4 parts. A fuzzifier converts knowledge into fuzzy knowledge or Membership Functions (MFs). The fuzzy rule base contains the relations between the input and output. The fuzzy illation method combines MFs with the management rules to derive the fuzzy output, and therefore the deffuzifier converts the fuzzy numbers back to a crisp worth. There are two reasons that symbolic logic systems are preferred: fuzzy systems are appropriate for unsure or approximate reasoning and that they permit higher cognitive process with calculable values underneath incomplete or unsure data. By using a fuzzy system to adaptively change the training parameters of the neural network in keeping with the MSE error, it is possible to cut back the chance of overshooting throughout the training method and facilitate the network to get out of an area minimum. There are four parameters accustomed to produce the principles for the symbolic logic management system; the relative error (RE), amendment in relative error (CRE), sign amendment in error (SC) and accumulative total of sign amendment in error (CSC).

L. Zhang et al. (2016) has given that deep multi-layer neural networks have several levels of non- dimensionality permitting them to succinctly represent extremely non- linear and extremely variable functions. The coaching section of deep neural network contains two major steps of parameter data format and fine standardization. The data format step is vital in deep learning. A stronger robust data format strategy might facilitate the neural network to converge to a good local minimum more efficiently. The fine standardization step permits to exactly adjust the parameters within the neural network in a much supervised way to enhance the discriminate ability of the ultimate feature.

In our proposed methodology convolutional algorithm has been used for detection and classification of cracks.

L. Zhang et al. (2016) have proved that Crack detection is an important application of neural networks. Steps for detection and classification of cracks were suggested by them were

a) Data preparation

The preparation of the data has to be done first for the process of crack detection.

b) Design and train the convolutional neural network

A deep learning algorithm could be designed to have many layers. Second step that has to take place is designing and training of the convolutional neural network.

c) Evaluate the performance of the convolutional neural network

After that the performance of the convolutional neural network has to be evaluated. The convolutional neural networks could be compared with the Support Vector Machine and other methods such as K nearest neighbor algorithm. The convolutional neural network requires less training and it has got the ability to detect complex relationships between the dependent and independent variables.

Motivation

The motivation to take up crack detection as the research area was because in the current scenario we did not have an appropriate maintenance policy for the safety of buildings and concrete surfaces and as a consequence the quality of the building degrades which in turn causes threat to the security of humans. In order to improve the safety and security of humans crack detection of concrete surfaces have been selected as the research area.

Crack detection in infrastructure building walls using Convolutional Deep Learning Algorithm.

Research Objectives

• a) To detect cracks in building walls using Deep Learning Algorithm.
• b) To detect the dimensions of the crack.
• c) To compare the convolutional algorithm with other algorithms and to determine which algorithm provides accurate results.

The first hypothesis of this research is to do a comparison on two of the efficient filtering methods and two of the best edge detection methods. They are average filter and median filter for filtering and Canny and Sobel for edge detection. In other words, the methods that could present a clear outline of the crack with less noise will be used in the later phases of crack detection.

The second hypothesis is to apply Convolutional Deep Learning Algorithm for the crack detection.

The third hypothesis is to compare the results of Convolutional Deep Learning algorithms with other algorithms such as Support Vector Machine and K Nearest Neighbor Algorithm to find out the accuracy.

Limitations of the Research

• Some independent small cracks cannot be identified using this method.
• Shadow noise and object influences are not considered in this study.
• The study was performed at the similar environmental conditions such as similar weather, existence of fog, hue of the concrete surface, the shape of structures which means that the environmental conditions were alike and the proposed algorithm needs to be evaluated in various fields of application.

This study provides a new crack detection method to detect cracks based on high resolution pictures. This method is more efficient and effective than the traditional method. The crack detection model is a fundamental process for the Visual Pattern Recognition (VPR) model.

This crack detection model could greatly reduce the computing cost for crack detection and it would save time and money while evaluating cracks in building walls and other concrete surfaces.

Challenges

Though Deep Learning algorithms achieve promising performance in multiple fields, there are many challenges still exist in this field. The two major challenges are,

a) Time complexity

b) Theoretical understanding.