Fog Removal. Its on Intelligent Surveillance. Fog removal from Camera 1800 words around that is fine.

1.Abstract

2.Introduction

3. Literature Review

4. Evaluation

5. Project Plan

6. References

Abstract

This proposal aims to review the high level tasks involved in object classification, specifically the classification of images of knives, spoons, and forks into their respective categories. The first major task is feature extraction; this involves translating features of the image such as colour and shape into their mathematical representation as vectors. The feature vectors can then be input into an image classifier which computes the probability that the image belongs to one category over another. Various types of features and image classifiers are evaluated.

Introduction

In order for a computer system to analyse and ultimately understand an image we have to first solve the problem of image classification. The objective of image classification is to categorise an image based on its content. An image classifier takes features of the image as input and uses statistical probability to surmise the category of the image. There is no one size fits all approach to image classification as different types of visual content have different statistical properties. For this reason it is important to select the correct image features and image classification algorithm in order to achieve a successful result.

Literature Review Feature Extraction

Feature selection is specific to the images that are to be classified. The number of features we select can impact the accuracy of the classification (Hughes, 1968). As knives, spoons, and forks are all similar in size and colour, differences in these features would be minor or non-existent. Shape is the most obvious difference between these items so focus will be put on extracting this features type.

Shape representation can generally be classified in one of two ways (Zhang and Lu, 2003). Contour-based methods extract feature information from only the boundary of the shape while region-based methods extract feature information from the entire shape region. Region-based methods are generally more accurate than contour-based methods as more information can be gathered from the entire region of the shape as opposed to only the boundary. Contour-based methods although being more common in literature, suffer from limitations including being more sensitive to noise and variations.

Moments are the statistical distribution of information of a region and thus are region-based features. Moments can describe different region features including area, centre of mass, and the angle of an axis (Prokop and Reeves, 1992). Moments are invariant when they are unaffected by changes to scale, translation, rotation, and reflection (Belkasim, 1991). The decision on which of these parameters are important is dependent on the classification problem and will ultimately determine the type of moment that is required.

Other types of moments include complex moments, rotational moments and orthogonal moments. Teague (1980) used Zernike orthogonal polynomials to create Zernike moments that could be used in pattern recognition. This type of moment has been studied extensively for use as an image feature. In terms of accuracy Zernike moments and pseudo-Zernike moments are preferable as an image feature for pattern recognition (Teh and Chin, 1988).

For the classification of knives, spoons, and forks, invariant moments, geometric moments and Zernike moments are all good candidates and will be evaluated and compared. Geometric moments have been shown to be a good representation of shapes with clear boundaries while Zernike moments can be used in more general use-cases and are less affected by noise than other types. As we will be classifying new or unseen images in our testing set, moment invariance is important. Zernike and geometric moments are superior to invariant moments in terms of performance (Khotanzad and Hong,1990).

For the purpose of comparison the moment features of the image, once extracted, need to be represented as a set of mathematical values. For this requirement a set of feature vectors will be created (Babu et al, 1997) which will contain the moment values of each image.

Image Classification

As is the case with image features there are multiple options available for the image classifier although many have been superseded by later generations that achieve better performance (a lower error rate). Image classifiers can be divided into two categories. The first, learning based or parametric classifiers such as the Artificial Neural Network (ANN) rely on a training period. Non-parametric classifiers such as Nearest Neighbour do not require any training time, can handle a large number of classes and avoid the over-fitting problem (Boiman et al, 2008).

The Nearest Neighbour classifier and many other non-parametric classifiers have been made obsolete primarily due to the performance difference. Multi-layer perceptron (Artificial neural network) based (parametric) classifiers outperform Nearest-Neighbour, Bayes and minimum-mean-distance (non-parametric) classifiers especially with noise (Alireza et al, 1990). Learning based classifiers including ANNs also outperform non-parametric classifiers with regards to error rate (Bottou et al, 1994). An issue with ANNs is that high levels of accuracy come at the expense of considerable training time. Although the classification accuracy of ANN can be very good the Support Vector Machine (SVM) is now generally thought of as being superior to the ANN with regards to generalisation accuracy in image classification (Burges, 1998) (Song, 2004). This means the SVM has higher levels of accuracy when classifying an unknown testing set.

The SVM achieves the higher generalisation accuracy because it uses an approach to learning that differs from that of the ANN. This approach is called Structural Risk Minimization (SRM) and has been shown to be superior to the Emperical Risk Minimization (ERM) approach of the ANN. The SVM is known to generalise well even in high dimensional spaces under small training sample conditions (Jonsson et al, 1999).

The SVM is a binary classifier; it will determine the hyperplane that best separates two classes of information from the infinite number of hyperplanes that could exist. The optimal separating hyperplane is the one that gives the largest margin between the two classes (Byun and Lee, 2003).

In the best case these two classes are linearly separable, however in most real world scenarios this is not the case. To remedy this, Boser et al (1992) created the non-linear SVM by adding a kernel function to the algorithm. This allows the data to be mapped to a higher-dimensional space in a way that then allows it to be linearly separable. (Cortes and Vapnik, 1995). This allows the SVM to be applicable to a wide range of image classification tasks.

As the SVM is a binary classifier it can only compare two classes of information. However in many cases we have more than two classes in which an image can be categorised. Rather than trying to turn the SVM into a multiclass classifier it is more practical to do several "one-against-one" classifications (Hsu and Lin, 2002). "One against one" applies pair-wise comparisons between classes while "One against the others" compares a given class with all the others put together (Chapelle et al, 1999).

Evaluation

The Weka data mining software from Waikato University will be used for the learning algorithm. This is mature software that has been used extensively in research.

The image features of the training set and the testing set will be extracted and converted into feature vectors, in a format that is supported by Weka. Feature vectors will be created for multiple types of moments as discussed in the literature review.

Image segmentation is an important aspect of moment feature extraction. As image segmentation is not in the scope of this project all images will be pre-processed manually so that the object to be classified is sitting on a white background. This will increase moment accuracy.

The training set is important with any learning algorithm. In order to avoid over-fitting, the correct amount of training data and feature representation is required. This will be determined during the project using trial and error based on the precision and recall of the results.

The support vector machine function in Weka will be used as the image classifier. In order to compare a test image to a knife, spoon, and fork, multiple binary classifications will be performed.

Measure

Recall and precision are the primary measures of the SVM and determine the proportion of classified images which are categorised correctly. Recall is expressed as true positives / (true positives + false negatives). Precision is expressed as true positives / (true positives + false positives).

Ultimately the accuracy of the image classifier against different moment training and test sets will determine the degree to which classification of knives, spoons and, forks is achievable.