A MATLAB-based Convolutional Neural Network Approach for Face Recognition System Article (PDF Available) January 2016 with 2,749 Reads How we measure 'reads'. Convolution is a formal mathematical operation, just as multiplication, addition, and integration. Addition takes two numbers and produces a third number, while convolution takes two signals and produces a third signal. Convolution is used in the mathematics of many fields, such as probability and statistics.

• Matlab Code For Newton's Method
• Matlab Convolution Of Images

Project Summary. Proposed a damage identification and localization technique for masonry structures using convolutional neural networks (based on AlexNet-8 and GoogLeNet-22 using Caffe).

Applied region-based convolutional neural networks (Faster-RCNN) to damage detection for masonry structures. Abstract Manual inspection (i.e., visual inspection and/or with professional equipment) is the most dominant approach to identifying and assessing superficial damage to masonry historic structures at present.

However, such method is inefficient, heavily relying on professionals and their experience. To overcome those limits, in this project, two deep learning-based damage detection techniques for masonry historic structures are proposed. One is using a sliding window-based deep learning method to identify and localize 3 damage categories (crack, efflorescence and spall) with the accuracy of 90.9%, the other is using a region proposal-based deep learning method to detect damaged brick units in quasi real-time (0.047 seconds per image). The advantage of deep learning method is that with sufficient training samples, machines are able to learn how to detect the damage automatically via training without using manual feature extraction techniques. Once the model training is complete, it can be used to detect damage directly.

Matlab Code For Newton's Method

The detection requires hardly any professional abilities or experience, and is hardly affected by the environment. Hence, high efficiency and accuracy can be realized. Introduction State-of-the-art structural health monitoring (SHM) techniques have been studied and applied to historic structures on account of their complexity, heterogeneity and long tradition of existence (1,2,3,4). Masonry has always been one of the key materials of historic structures, mainly in the form of walls which carry loads and provide security. The safety of masonry structures is of vital importance in the maintenance of historic architecture (5,6).

As structures gradually deteriorate, they are likely to be subjected to various forms of damage, mainly resulting in the reduction of strength, stiffness and integrity. Such damage always reflects on the surface of structures, such as crack, efflorescence, spall, and so forth. Hence, it is of great significance to identify and localize the superficial damage quickly and efficiently. The work can be regarded as guidance for structural repair, maintenance and management afterwards (7).

The most dominant superficial damage detection method in actual practice is in-situ visual inspection supplemented by professional equipment (8,9,10,11). Such method seems easy to operate, but requires extremely high professional inspection and experience. For instance, in terms of crack detection, the operators must be capable of fully understanding the properties of the inspected cracks, and determine whether repairs are needed. Large-scale inspection using this method is a time and labor-intensive process (12), so it can only be implemented via sampling detection and periodic inspection.

Although manual visual inspection can achieve reasonable performance in most scenarios, sometimes it is unreliable as serious damage is not detected (13). Hence, such method may not provide necessary guidance for maintenance in historic structures, resulting in a substantial increase in maintenance costs. One solution to those problems is using sensors-based SHM systems.

Matlab Convolution Of Images

However, collecting data by various sensors and monitoring every historic structures in real-time require a huge expense. Moreover, SHM systems that integrated and installed into the historic structures also require professionals to operate and the distributed devices are difficult to implement on large-scale structures (14). Therefore, it may not be a cost-effective choice. In recent years, plenty of research on vision-based SHM techniques has been carried out, aiming to replace the conventional manual inspection (15,16,17,18,19,20). Since professionals are able to draw conclusions via only visual inspection, it is possible to use computers to detect the damage automatically based on digital image processing. However, such method requires manual feature extraction. The images of the structures are sensitive to noise such as stains, shadows and non-uniform lighting condition.

Hence, the robustness and adaptability of digital image processing is not sufficient. A number of studies have introduced machine learning techniques to improve the accuracy. Kabir (21) proposed a SHM technique using grey level co-occurrence matrix (GLCM) and neural network classifier to detect alkali-aggregate reaction (AAR) damage. Moon and Kim (22) detected cracks with reasonable accuracy based on image processing and backpropagation (BP) neural networks. O’Byrne et al. (23) proposed an enhanced texture segmentation technique for structural damage detection and classification based on GLCM and support vector machines (SVMs).

Plevris and Asteris (24) approximated the failure surface of masonry structures under biaxial stress using neural networks. (25) proposed a crack-defragmentation technique based on dilation transform and thinning transform, and used neural network classifier to identify cracks afterwards. However, these methods based on shallow neural networks and SVMs still use image processing techniques such as edge detection. Hence, the robustness and adaptability are not significantly improved. Is there any vision-based detection approaches require no image processing technique? The answer is deep learning (26).

Compared with fully connected BP neural networks-based shallow learning, convolutional neural networks (CNNs)-based deep learning requires no manual feature extraction. The classifier is trained via automatic machine learning instead.

The reason deep learning can be “deep” is that the convolution and pooling techniques in CNNs are able to significantly reduce parameters, making very deep networks possible. CNNs are derived from “receptive field” proposed by Hubel and Wiesel (27) who are inspired by cat’s visual cortex. With the development of BP techniques, LeCun (28) first proposed a CNN for document recognition, which is the famous LeNet. However, CNNs were not more efficient and popular than normal BP neural networks owing to the low-performance hardware and insufficient training data. CNNs owe much of their performance to the realization of big data and parallel computing on graphics processing units (GPUs) (29). In the Large Scale Visual Recognition Challenge 2012 (LSVRC2012) (30), Krizhevsky et al.

(31) first used a CNN model and got the champion. The model is called AlexNet which achieved the top-5 test error rate of 15.3% (Top-5 error rate of the best model in LSVRC2011 is 25.8%). From 2012, all champion models of LSVRC are CNNs, such as ZFNet (32) in 2013, GoogLeNet (33) in 2014, and ResNet (34) in 2015. CNN models are going deeper and deeper to improve the performance, achieving very high accuracy, robustness and adaptability in image classification. Due to the high performance of deep learning in image classification, it is really promising to apply them to structural damage detection. However, not much research on deep learning-based structural damage detection has been proposed.

Ouellette et al. (35) used a genetic algorithm-based CNN to detect cracks with not very high accuracy.

Makantasis et al. (36) used a CNN with only two convolution layers and a fully connected layer in tunnel crack inspection. Protopapadakis et al. (37) combined CNNs with 3D modeling techniques and applied them to detect concrete defects. (38) proposed a complete strategy for crack detection using a CNN and a sliding window technique with high accuracy. All these studies above only use CNNs to classify two categories (i.e., intact and crack), and the CNN models are not deep (no deeper than AlexNet).

The state-of-the-art region proposal-based deep learning methods have never been used for damage detection either. Hence, two contributions are aimed to be made in this project.

First, more than one type of damage detection is first realized, and CNN models in this project are going deeper (the deepest CNN in this project is based on GoogLeNet with 22 layers). Second, a region proposal-based deep learning method for damage detection is also first utilized. 1: Schematic diagram of image classification for a brick unit Fig. 2: Schematic diagram of CNN-based damage identification and classification for a brick unit Fig. 3: Comparison of fully connected mode and locally connected mode: (a) fully connected mode, and (b) locally connected mode Fig. 4: Training samples of the 4 categories: (a) intact, (b) spall, (c) crack, and (d) efflorescence for CNN Fig. 5: Training sample for Faster-RCNN Fig.

6: Results of training using AlexNet-8: (a) train loss, and (b) validation accuracy Fig. 7: Results of training using GoogLeNet-22: (a) train loss, and (b) validation accuracy Fig. 8: Sliding window strategy Fig.

9: Test results based on GoogLeNet-22 together with sliding window algorithm Fig. 10: Test results based on Faster-RCNN.