Classification of Rusty and Non-Rusty Images: A Machine Learning Approach

Introduction

Corrosion is a natural process that occurs in metals in contact with humidity and pollution (Sedriks, 1996a). Due to corrosion buildings and bridges can collapse, oil pipelines break and many more (Acosta et al., 2014a). Corrosion of metallic materials has been an inevitable part of the human experience. It can cause huge damage in industrial areas. While the oxidation of iron which is called rust, is the most easily identified form of corrosion, this oxidation process represents only a fraction of material losses (Fontana, 2005). Therefore, it is crucial to detect it in the early stages so that it can be prevented without causing severe damages. For this purpose industries are using modern techniques which includes human involvement in making final review of the rusted parts. This includes human error in the decision making.

The human error involved can be eliminated with the proposed model for rust detection which implements Digital Image Processing technique that automates the visual inspection through statistical features of the rusted areas of the parts (Sedriks, 1996a). A database of the images containing rust is used for analysis purpose. The need for thousands of images for training purpose motivate us to generate multiple images using Perlin noise (Acosta et al., 2014a).

Proposed model involves three stages namely acquisition, processing and classification. Acquisition stage involves capturing images using physical devices like cameras and sensors. Processing involves filtering, feature extraction with the help of which we can per- form classification. Classification is used to classify the images into rusty and non-rusty classes. For improving accuracy different types of classifiers are used like Bayesian classifier, Support Vector Machine.

• •

  Motivation

As discussed earlier, Corrosion creates huge damage to the industries and there is need to detect it at earlier stages. Though, modern techniques are available to detect the rust, they are introducing human error in the detection process. Hence the idea of removing that human error motivate us to design this detector model and improve the accuracy of the classifier (Fontana, 2005).

• •

  Challenges

The main challenges are Camera Quality, Presence of Noise in the Image, Presence of various texture and color of rust (Acosta et al., 2014a). In image acquisition stage we need to capture images with the help of good quality camera, otherwise it will be more challenging to handle the captured images in further stages (Reinhard et al., 2010). Second challenge is presence of noise in the images. Though the camera for capturing used is of good quality, sometimes the images will have noise which needs to be removed by filtering process. The third challenge is presence of various texture and color of rust which increases the difficulty level in detecting it (Acosta et al., 2014b). Hence the need to train our model to detect various textures and color of rust.

• •

  Goals and Objectives

In the recent era automated and correct detection of rusty and non-rusty area plays very important role in the industry. However, the online databases available are not sufficient in size which leads to overfitting of the model. Hence, the ultimate aim of the paper is to generate the sizable dataset, find the best suitable features and classifier which can be used for automation in the devices used to identify rusty and non-rusty images.