Choong Yeun Liong

Review on image enhancement methods of old manuscript with the damaged background

Quite often that old documents are suffering from background damage. Examples of background damages are varying contrast, ancient document age and the documents have degraded over time due to storage conditions and the quality of the written parchment. These have damaged background for example such as: have varying contrast, smudges, dirty, presence of seeping ink from the other side of the document, uneven background. In order to make them readable, image processing offers a selection of approaches. The aim of this paper is to provide comprehensive review methods to enhance old document images with damaging background. Three kinds of enhancement methods are: (a) Image enhancement methods using binarization method or thresholding method, (b) Image enhancement methods using binarization method or thresholding method and other methods, (c) Image enhancement methods using other methods only. As conclusion, the second method has becoming more popular and has a great potential to improve in future.

Firearm identification using numerical features of centre firing pin impression image

There are many crime cases such as murders or robberies which frequently involve firearms, especially pistols. The centre firing pin impression image on a cartridge case is one of the important clues for firearms identification. In this study, a total of 16 features of geometric moments up to the sixth order were extracted from centre of firing pin impression images. A total of five pistols of the Parabellum Vector SPI 9mm model, made in South Africa were used. The pistols were labelled as Pistol A, Pistol B, Pistol C, Pistol D, and Pistol E. A total of 747 bullets have been fired from the five pistols. Under preliminary analysis, Pearson correlation coefficients between all pairs of features showed the features were significant and highly correlated among the features. This problematic features were solved by dividing the features into subgroups of variables based on similar characteristics under principle component analysis. The features that highly correlated were combined into meaningful components or factors. Discriminant analysis was applied to identify the types of pistols used based on the factors obtained. Classification results using cross-validation under discriminant analysis showed that 75.4% of the images were correctly classified according to the pistols used. The results of the study had shown a significant contribution towards Royal Malaysian Police Force in handling crime cases which involve firearms in more systematic manner.

Firearm recognition based on whole firing pin impression image via backpropagation neural network

Firearms identification is a vital aim of firearm analysis. The firing pin impression image on a cartridge case from a fired bullet is one of the most significant clues in firearms identification. In this study, a set of data which focused on selected 6 features of firing pin impression images before an entirety of five different pistols of South African made; the Parabellum Vector SPI 9mm model, were used. The numerical features are geometric moments of whole image computed from a total of 747 cartridge case images. Under pattern recognition theory, the supervised features of firing pin impression images were then trained and validated using a two-layer backpropagation neural network (BPNN) design with computed hidden layers. A two-layer 6–7–5 connections BPNN of sigmoid/linear transfer functions with ‘trainlm’ algorithm was found to yield the best classification result using cross-validation, where 96% of the images were correctly classified according to the pistols used. Moreover, the network was trained under very small mean-square error (MSE=0.01). This means that neural network method is capable to learn and validate well the numerical features of whole firing pin impression with high precision and fast classification results.

Invariant Features from the Trace Transform for Jawi Character Recognition

The Trace transform, a generalization of the Radon transform, allows one to construct image features that are invariant to a chosen group of image transformations. It consists of tracing an image with straight lines along which certain functionals of the image function are calculated. It can be useful to construct invariant features to rotation, translation and scaling of the image. In this paper, we demonstrate the usefulness of the features in classification of Jawi character images using multilayer perceptron neural networks. We compare the result of character recognition with those obtained by using features based on affine moment invariants.

Effects of scatter-correction pre-processing methods and spectral derivative algorithms on forensic classification of paper

Infrared (IR) spectral data are always influenced by undesired random and systematic variations. As such, pre-processing of spectral data is normally required before chemometric modeling. Two most widely used pre-processing techniques, i.e. scatter-correction methods and spectral derivatives, were used to pre-process 150 IR spectral data of paper. The algorithms investigated in this preliminary study are Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), Savitzky-Golay (SG) and Gap-Segment (GS). The visual examination of the clustering among three studied varieties of paper, i.e. IK Yellow, One Paper and Save Pack, is accomplished via Principal Component Analysis (PCA). Overall, separation of the three varieties of paper is greatly enhanced after pre-processing. The most significant improvement is obtained with pre-processing via 1st derivative using SG algorithms.

The effects of column-wise manipulations on accuracy of classical classifiers with high-dimensional spectral data

Column-wise manipulations (CWM), a group of data pre-processing (DP) techniques composed of mean-centering, Pareto scaling (PS), variance scaling and auto-scaling; are often applied individually or in combination. It has been applied like a norm without thoughtful considerations partly attributed to its simplicity and ease of applications. Theoretically, all variables in IR spectrum are measured on the same scale and seldom have different means and as such rarely require CWM as compared to normalization. This preliminary paper aims to investigate the real needs of each aforementioned CWM in infrared (IR) spectroscopic dataset that is derived from white copy paper. The untreated and pre-processed IR data is then processed with Principal Component Analysis plus Linear Discriminant Analysis (PCA-DA). The impact of CWM on test accuracy of the different PCA-DA models is then compared according to different IR wavenumber intervals. Error of the predictive models is determined via nonparametric bootstrap. Result...

Forensic differentiation of paper by ATR-FTIR spectroscopy technique and partial least-squares-discriminant analysis (PLS-DA)

The infrared (IR) spectra of different white copy paper types tend to appear indifferent. Discrimination between white copy papers could lead to the solution of forgery cases. In this preliminary study, three varieties of white paper were purchased from local stationery shops in Kuala Lumpur, Malaysia. The papers were classified according to their manufacturers using Partial Least Squares-Discriminant Analysis (PLS-DA) and sparse PLS-DA (sPLS-DA) models. The error rate for the two models on the training and the test data sets were estimated and compared. Results show that the performance of the two models are comparable. By including the first six latent variables in both models, classification accuracy as high as 100% can be achieved.

Genetic algorithms for wavenumber selection in forensic differentiation of paper by linear discriminant analysis

Selection of the most significant variables, i.e. the wavenumber, from an infrared (IR) spectrum is always difficult to be achieved. In this preliminary paper, the feasibility of genetic algorithms (GA) in identifying most informative wavenumbers from 150 IR spectra of papers was investigated. The list of selected wavenumbers was then employed in Linear Discriminant Analysis (LDA). GA procedure was repeated 30 times to get different lists of variables. Then the performances of LDA models were estimated via leave-one-out cross-validation. A total of six to eight wavenumbers were identified to be valuable variables in the GA procedures. All the 30 LDA models achieve correct classification rates between 97.3% to 100.0%. Therefore the GA-LDA model could be a suitable tool for differentiating white papers that appeared to be highly similar in their IR fingerprints.

Comparison of several variants of principal component analysis (PCA) on forensic analysis of paper based on IR spectrum

Principal Component Analysis (PCA) is a commonly used unsupervised exploratory analysis technique. It is also frequently used in dimensionality reduction. This preliminary paper investigates the feasibility of three variants of PCA, i.e. independent PCA (iPCA), sparse PCA (sPCA), and sparse independent PCA (siPCA) on forensic classification of paper based on their IR spectral data. After that, Linear Discriminant Analysis (LDA) models were built using the Principal Components (PCs) produced by the PCA and the three aforementioned variants. The performances of all these four LDA models, i.e. PCA-DA, iPCA-DA, sPCA-DA and siPCA-DA, were evaluated via leave-one-out cross-validation on the data set. The results obtained show that iPCA-DA and siPCA-DA are the most effective models with 100.0% classification accuracy. Then, the effectiveness of siPCA and iPCA models was evaluated based on posterior probability used for predictions of class membership that were derived from leave-one-out cross-validation. As a conclusion, siPCA is identified as the best classification model.

Adaptive binarization method for enhancing ancient malay manuscript images

In order to transform ancient Malay manuscript images to be cleaner and more readable, enhancement must be performed as the images have different qualities due to uneven background, ink bleed, or ink bleed and expansion of spots. The proposed method for image improvement in this experiment consists of several stages, which are Local Adaptive Equalization, Image Intensity Values, K-Means Clustering, Adaptive Thresholding, and Median Filtering. The proposed method produces an adaptive binarization image. We tested the proposed method on eleven ancient Malay manuscript images. The proposed method has the smallest average value of Relative Foreground Area Error compared to the other state of the art methods. At the same time, the proposed method have produced the better results and better readability compared to the other methods.