Image Classification using SVM Techniques

Study for Image Classification using the techniques under SVM for computer vision

Project Dissertation

Student name:

Student ID:

Chapter 1: Introduction

Background

With the advancements in Remote Sensingtechnology over the past decade, Earth Observation data has become more readily available in poor countries. Sensor and information infrastructure enhancements and platform innovations are all included in this category of services. Traditional data collection methods, which will be addressed in greater detail below, are also less favourable for agricultural and environmental applications, owing to the high-quality data availability, the frequency with which data is collected, and the extent to which data is covered. Because of the increased accessibility of RS picture archives, LULC mapping has emerged as the most extensively utilised RS data application for a wide range of environmental research projects, particularly in the agriculture and forestry industries and other purposes. In response to the growing demand for LULC mapping applications and the necessity of updating existing maps, researchers are investigating the development of new image classification algorithms in a variety of land management industries, including agriculture, forestry, and forestry products. Novel remote sensing picture categorization algorithms are being studied to address land management concerns on a regional, local, and global scale.

For a multitude of reasons, categorization is difficult to accomplish. Access to massive volumes of remote sensing data and complex scenarios with little training data are examples of such challenges. Because of how well the system performs with remote sensing image classification and how many training samples are required to train it, input parameters, pre-processing procedures, target classes, supplementary data, and accuracy are all factors that influence the system's efficiency and computation cost.When attempting to increase the final classification map accuracy, it is critical to keep the following factors in mind. When it comes to evaluating fundamental accuracy, it is likely that the most frequently used and trusted methodologies used to report the correctness of the thematic maps, such as the Overall Correctness and the Kappa coefficient of agreement, will be utilised. By including distinct training and data validation sets, it is possible to check the classification system's correctness. Using this information, it will be possible to compare and contrast the various categorization techniques available (Sheykhmousa, et. al., 2020).

Many review papers have been written in response to the implementation and development of novel categorization algorithms, which have been implemented and developed over time. Researchers have been studying RS classification algorithms for more than a decade to gain a better understanding of the fundamental wide range characteristics of methodologies and procedures, as well as their limitations and weaknesses. Further research has revealed that these algorithms perform admirably under a wide range of situations, making them applicable to a wide range of applications.There are several examples of deep learning review papers, aside from those that compare the capabilities of architectures for deep learning with the Support Vector Machines for image classification, that is focused on hyperspectral data classification using the deep learning approaches. Thus, this research will focus in-depth discussion of the tactics for image classification using SVM (Song, et. al., 2018).

The technique has been applied in picture categorization research to develop a plethora of different computer vision techniques. Things like moments of first-order and hand-made features are examples of what we're talking about. This new thing has been noted by many people. We were able to start the development process in its entirety by utilizing moments of first order and features that are dependent on the grey level of the environment. Participants in the study were given a list of distinct textural characteristics to consider. It is more accurate than other methods of texture classification since their method is based on grey-level dependent statistics rather than a binary classification system. Structure information about textures is being created to aid in the grouping of textures.

According to another study, moments of the first order can be utilized to obtain information about the texture of photographs by analyzing them. This information can be used in conjunction with content-based photo search. Second-order moments can be less stable in terms of scale and rotation than first-order moments, and this is especially true for rotational moments. When it comes to rotation and size, it is invariant. As a result, Gabor filters have been developed and are currently being employed in texture categorization to deal with the issues that they present. Because machine learning is not appropriate for all situations, supervised learning has been made accessible as an alternative.

It demonstrates that utilizing first-order moment as it is not possible to consistently distinguish between two textures that have the same contrast level. This method is also ineffective when dealing with concepts and ideas that are difficult to comprehend and absorb in their whole. According to a new study, local binary patterns can be utilized to obtain more precise information about the local environment while also obtaining more data. LBP is useful by all prior computer vision algorithms since it is basic and straightforward to utilize (Shaha, et. al., 2018).

Although there is no direct information provided, it does not important because there are many distinct LBP versions that can be used to fit your requirements. In a variety of investigations, researchers have discovered a means to obtain information about a location's lighting local features which is invariant from a diverse range of data sources by employing local ternary patterns. Several studies have concluded that the CBIR should have access to a database of local service providers (Cao, et. al., 2020). This is something that should be done. There are twelve different orientations that this design takes and conveys information, making it more durable than other patterns.

When using hyperspectral imaging, it is possible to detect faults in some fruits, such as apples and blueberries (HSI). Over a decade, this technology has been in use. It can perceive things that other foods do not have in the same way. The techniques of image processing and spectrum analysis are combined in this strategy. It is difficult to employ hyperspectral imaging (HSI) in the field because it takes a long time for acquiring and examine HD hyperspectral images, and then figure out what they signify. This is because it takes a long time to get and examine high-dimensional hyperspectral images (Yang, et. al., 2018).

According to a new study published in the Journal of Food Science, shortwave infrared HSI can be used to detect apple bruising in the field. It is capable of doing so with 98 per cent accuracy and in less than 200 milliseconds of each apple. To figure out which apple was which, apples from three different varieties were used: Jonagold, Joly Red, and Kanzi. This was done for the algorithm to be able to determine which apple was which. When it comes to all three types of apples, the computer had a 94.4 per cent accuracy rating in terms of prediction. As a result, high-speed imaging (HSI) devices were not widely employed in the packaging business since they were expensive to purchase at the time of purchase. A multispectral imaging system, on the other hand, can be utilized for online detection in place of a high-speed imaging system, which only captures photos at specific wavelengths.

When it came to detecting citrus canker, a technique known as MSI was developed. Its key wavelengths were 730 nm as well as 830 nm. It was capable of detecting five different fruits in a single second. The system was able to recognize five fruits every second by using the wavelengths 730 and 830 nm as its primary wavelengths. It was hard to discover flaws in bi-colored fruits such as citrus and bananas then it was in mono-coloured fruits such as oranges and apples (such as apples). There were certain places on the fruit that seemed identical to one another. For the time being, an MSI system is unable for detecting apple problems online due to the poor quality of pictures obtained using the beam splitting approach and the resemblance between stems/calyxes and defects MSI systems may not be widely used in the commercial sector since they are difficult to operate, expensive, and enormous (Fan, et. al., 2020).

Research Aim

The aim of this research is to present the techniques under SVM acquired from the different research study to analyze their effectiveness for the image classification.

Research Objective

The research objective are-

• Presenting historical development of SVM techniques for image classification

• Presenting the working of techniques under SVM for image classification

• Presenting how accurate the showcased technique are for image classification based on the result gained from different research study

Research Questions

• Question 1 – What are the historical development of SVM techniques for image classification?

• Question 2 – How this techniques works under SVM for image classification?

• Question 3 – How accurate these techniques gives result for image classification?

Research Significance

It is possible to improve classification accuracy by using strategies that are hybrid in nature. Based on the situation, hybrid techniques might be superior to single procedures in some cases. Classifying SVM data can be done in several ways, and thus the presented study focuses on how these different methods may be used to classify data for different reasons.

Research Map

The following points represent the aim of the proposed study-

• The first section will include historical developments for the chosen topic.

• A thorough explanation of the research methodology employed in the presented study, which can be found in the second section of the article.

• Having provided a brief overview of the SVM and its related supporting techniques will be presented in the third section.

• The fourth section will go into greater detail regarding SVM and how it interacts with other concepts. Images can be categorized in greater detail than in the previous section, which relies on the usage of SVM in the fourth section to further refine the classification.

• The fifthsection willdelve deeper into the subject as this section is followed by the last chapter, which serves as a continuation of the previous chapter.

• Lastly, concluding with the overall explanation about discussed technologies in the presented study.

References

Sheykhmousa, M., Mahdianpari, M., Ghanbari, H., Mohammadimanesh, F., Ghamisi, P. and Homayouni, S., 2020. Support vector machine versus random forest for remote sensing image classification: A meta-analysis and systematic review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13, pp.6308-6325.

Song, W., Li, S., Fang, L. and Lu, T., 2018. Hyperspectral image classification with deep feature fusion network. IEEE Transactions on Geoscience and Remote Sensing, 56(6), pp.3173-3184.

Shaha, M. and Pawar, M., 2018, March. Transfer learning for image classification. In 2018 second international conference on electronics, communication and aerospace technology (ICECA) (pp. 656-660). IEEE.

Yang, X., Ye, Y., Li, X., Lau, R.Y., Zhang, X. and Huang, X., 2018. Hyperspectral image classification with deep learning models. IEEE Transactions on Geoscience and Remote Sensing, 56(9), pp.5408-5423.

Fan, S., Li, J., Zhang, Y., Tian, X., Wang, Q., He, X., Zhang, C. and Huang, W., 2020. On line detection of defective apples using computer vision system combined with deep learning methods. Journal of Food Engineering, 286, p.110102.

Cao, X., Yao, J., Xu, Z. and Meng, D., 2020. Hyperspectral image classification with convolutional neural network and active learning. IEEE Transactions on Geoscience and Remote Sensing, 58(7), pp.4604-4616.