ALL PuBlications

Chit Su Hlaing  | Sai Maung Maung Zaw

Abstract

We introduce a set of statistical features and propose the SIFT texture feature’s descriptor model on statistical image processing. The proposed feature is applied to plant disease classification with PlantVillage image dataset. The input is plant leaf image taken by phone camera whereas the output is the plant disease name. The input image is preprocessed to remove background. The SIFT features are extracted from the preprocessed image. As a main contribution, the extracted SIFT features are model by Generalized Extreme Value (GEV) Distribution to represent an image information in a small number of dimensions. We focus on the statistical feature and model-based texture features to minimize the computational time and complexity of phone image processing. The propose features aim to be significantly reduced in computational time for plant disease recognition for mobile phone. The experimental result shows that the proposed features can compare with other previous statistical features and can also distinguish between six tomato diseases, including Leaf Mold, Septoria Leaf Spot, Two Spotted Spider Mite, Late Blight, Bacterial Spot and Target Spot.

Date: 18th – 20th December, 2017 |18th International Conference on Parallel and Distributed Computing, Applications and Technologies (PDCAT) | Weblink

Chit Su Hlaing  | Sai Maung Maung Zaw

Abstract

The current focus of our research is to detect and classify the plant disease in agricultural domain, by implementing image processing techniques. We aim to propose an innovative set of statistical texture features for classification of plant diseases images of leaves. The input images are taken by various mobile cameras. The Scale-invariant feature transform (SIFT) features used as texture feature and it is invariant to scaling, rotation, noise and illumination. But the exact mathematical model of SIFT texture descriptor is too complex and take high computing time in training and classification. The model-based statistical features are calculated from SIFT descriptor to represent the features of an image in a small number of dimensions. We derive texture information probability density function called Generalized Pareto Distributions from SIFT texture feature. The main focus of our proposed feature is to reduce computational cost of mobile devices. In our experiment, 10-Fold cross validation with SVM classifiers are applied to show that our experiment has no data bias and exclude theoretically derived values.

Date:  24th – 27th October, 2017 | IEEE 6th Global Conference on Consumer Electronics (GCCE 2017) | Weblink

Chit Su Hlaing  | Sai Maung Maung Zaw

Abstract

Plant disease classification is essential for food productivity and disease diagnosis in agricultural domain. The probability distribution and statistical properties are essential in image processing to define the features of typical image. The general usage of (Scale Invariant Feature Transform) SIFT has local feature extraction and global feature extraction (bag-Of-Features approach) for classification, and its classification result for unknown data also depends on code book (global feature) generation. Instead of using bag-Of- Feature approach, we proposed to apply Beta probability distribution model for SIFT to be directly represent the image information and then formed SIFT-Beta. The color statistics feature is extracted from RGB color space and then combines with SIFTBeta to produce proposed features. The proposed feature is applied in Support Vector Machine classifier. The classifier  is trained for seven labels of tomato with six diseases and healthy.

Date:  27th February, 2019 | Seventeenth International Conference On Computer Applications (ICCA 2019) | Weblink

Chit Su Hlaing  | Sai Maung Maung Zaw

Abstract

Plant disease classification has been associated with the production of essential food crops and human society. In this paper, we classify tomato plant disease using two different features: texture and color. For a texture feature, we extract statistical texture information (shape, scale and location) of an image from Scale invariant Feature Transform (SIFT) feature. As a main contribution, a new approach is introduced to model the Scale Invariant Feature Transform (SIFT) texture feature by Johnson SB distribution for statistical texture information of an image. The moment method is used to estimate the parameters of Johnson SB distribution. The mathematical representation of SIFT feature is matrix representation and too complex to be applied in image classification. Therefore, we propose a new statistical feature to represent the image in few numbers of dimensions. For a color feature, we extract statistical color information of an image from RGB color channel. The color statistics feature is the combination of mean, standard deviation and moments from degree three to five for each RGB color channel. Our proposed feature is a combination of statistical texture and color features to classify tomato plant disease. The experimental performance on PlantVillage database is compared with state-of-art feature vectors to highlight the advantages of the proposed feature.

Date:  6th -8th  June, 2018 | 17th IEEE/ACIS International Conference on Computer and Information Science, ICIS 2018, Singapore,  IEEE Computer Society 2018, ISBN 978-1-5386-5892-5 | Weblink

Aye Min  |  Nu War

Abstract

The detection of brain cancer without human interfering is a major problem in the domain of medicinal image processing. The segmentation of brain images of MRI is a technique used as a first step to extract different characteristics of these images for analysis, appreciative and understanding. The main function of brain segmentation by MRI is to detect the type of brain abnormality. Many segmentation techniques are proposed in the literature. In this comparative paper, we will discuss the behaviors of tested segmentation methods. Otsu thresholding, Region growing, Particle swarm optimization and Interactive graph cut segmentation methods are analyzed and compared in this paper. After segmented with these methods, the morphological operation is used to get exact shape and size of tumors. As a benchmark dataset, BRATS dataset is used to test segmentation results.

Aye Min  | Zin Mar Kyu

Abstract

With advanced imaging techniques, Magnetic Resonance Imaging (MRI) plays an important role in medical environments to create high quality images contained in the human organs. In the processing of medical images, medical images are coordinated by different types of noise. It is very important to acquire accurate images and observe specific applications with precision. Currently, eliminating noise from medical images is a very difficult problem in the field of medical image processing. In this document, three types of noise, Gaussian noise, and salt & pepper noise, uniform noise are tested and the tested variances of Gaussian noise and uniform noise are 0.02 and 10 respectively. We analyze the kernel value or the window size of the medium filter and the Wiener filter. All experimental results are tested on MRI images of the BRATS data set, the DICOM data set and TCIA data set. MRI brain images are obtained from the BRATS data set and the DICOM data set, the MRI bone images are obtained from the TCIA data set. The quality of the output image is measured by statistical measurements, such as the peak signal noise ratio ( PSNR) and the root mean square error (RMSE).

Aye Min  | Zin Mar Kyu

Abstract

Brain tumor is the abnormal growth of cancerous cells in Brain. The development of automated methods for segmenting brain tumors remains one of the most difficult tasks in medical data processing. Accurate segmentation can improve diagnosis, such as evaluating tumor volume. However, manual segmentation in magnetic resonance data is a laborious task. The main problem to detect brain tumors is less precise to determine the area of the tumor and determine the segmentation accuracy of the tumor. The system proposed the fusion based results binding for MRI image enhancement and combination of adaptive K-means clustering and morphological operation for tumor segmentation. BRATS multimodal images of brain tumor Segmentation Benchmark dataset was used in experiment testing.

Aye Min  | Zin Mar Kyu

Abstract

Brain tumor is the abnormal growth of cancerous cells in Brain. In medical field, segmentation of brain regions and detection of brain tumor are very challenging tasks because of its complex structure. Magnetic resonance imaging (MRI) provides the detailed information about brain anatomy. Proper brain tumor segmentation using MR brain images helps in identifying exact size and shape of Brain tumor, this intern helps in diagnosis and treatment of brain tumor. However, manual segmentation in magnetic resonance data is a time-consuming task and is still being difficult to detect brain tumor area in MRI. The main challenges of brain tumor detection are less of accuracy to detect tumor area and to segment the tumor area. The system proposed the results fusion method for image enhancement and combination of adaptive k-means clustering and morphological operation for tumor segmentation. All of the experimental results will be tested on BRATS multimodal images of brain tumor Segmentation Benchmark dataset.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

This paper presents a system for Myanmar text extraction and recognition from warning signboard images taken by a mobile phone camera. Camera captured natural images have numerous difficulties compared to the traditional scanned documents. Common problems for camera captured text extraction are variations in font style, size, color orientation, illumination condition as well as the complex background. In this system, color enhancement process is performed to distinguish the foreground text and background color. Color enhanced images are converted into binary images using color threshold range. The detected non-text objects are removed as clearly as possible using width, high, aspect ratio and object region area threshold. In the segmentation process, horizontal projection profile, vertical projection profile and bounding box are used for line segmentation and character segmentation. To recognize the above segmented Myanmar characters, blocking based pixel count and eight-direction chain codes features are proposed. In this system, characters are classified by feature based approach of template matching method by using the proposed features. In this paper, dynamic blocking based pixel count, eight-direction chain codes features and geographic features are used to correctly recognize Myanmar characters.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

This paper presents a very simple and efficient method for the text extraction and recognition of the Myanmar text from color natural signboard images taken by a mobile phone camera. Text extraction, line segmentation, character segmentation and recognition are the important steps in text understanding from natural signboard images. In this system, the color enhancement is firstly processed to overcome various illumination conditions. Background noises on the binary images are removed by four filtering features such as color threshold based filtering, aspect ratio based filtering, boundary based filtering and region area based filtering. After removing the noise, line segmentation and character segmentation are done. Horizontal projection profile is used for line segmentation and vertical projection profile and bounding box methods are used to segment the characters. These connected component characters are recognized by using 4×4 blocks based pixel density and total chain codes, 4-rows based pixel density, 4-columns based pixel density and count of eight directions chain code on the whole character image and on each block of character image. This system is investigated by feature based approach of template matching, and 83.15% character recognition accuracy is achieved on 2854 correctly extracted characters from 150 camera-captured Myanmar warning signboards.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

This paper publicizes the character segmentation and recognition of the Myanmar warning text signboard images taken by a mobile phone camera in natural scene. In this system, two templates are created. The first template that contains both connected pixel words and characters are used for character segmentation and the second template that contains only the connected pixel characters are used for character classification. Color enhancement process is first performed to extract the text regions. By preprocessing the color enhancement, the system can overcome the some illumination conditions. To remove the background noises on the binary images, color threshold based filtering, aspect ratio based filtering, boundary based filtering and region area
based filtering techniques are used. As a next step, line segmentation and character segmentation are done. Line segmentation is performed using horizontal projection profile and character segmentation is done using vertical projection profile and bounding box methods. In the character segmentation process, template matching method is used by training connected pixel words. These connected component characters are recognized using 4 × 4 blocks based pixel density and total chain codes, four rowsbased pixel density, four columns-based pixel density and count of eight directions chain code on the whole character image and on each block of character image. This system is investigated by feature-based approach of template matching on 160 cameracaptured Myanmar warning signboards.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

Character recognition is the process of converting a text image file into editable and searchable text file. Feature Extraction is the heart of any character recognition system. The character recognition rate may be low or high depending on the extracted features. In the proposed paper, 25 features for one character are used in character recognition. Basically, there are three steps of character recognition such as character segmentation, feature extraction and classification. In segmentation step, horizontal cropping method is used for line segmentation and vertical cropping method is used for character segmentation. In the Feature extraction step, features are extracted in two ways. The first way is that the 8 features are extracted from the entire input character using eight direction chain code frequency extraction. The second way is that the input character is divided into 16 blocks. For each block, although 8 feature values are obtained through eight-direction chain code frequency extraction method, we define the sum of these 8 feature values as a feature for one block. Therefore, 16 features are extracted from that 16 blocks in the second way. We use the number of holes feature to cluster the similar characters. We can recognize the almost Myanmar common characters with various font sizes by using these features. All these 25 features are used in both training part and testing part. In the classification step, the characters are classified by matching the all features of input character with already trained features of characters.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

In any country, warning text is described on the signboards or wall papers to follow by everybody. This paper present Myanmar character recognition from various warning text signboards using block based pixel count and eight-directions chain code. Character recognition is the process of converting a printed or typewritten or handwritten text image file into editable and searchable text file. In this system, the characters on the warning signboard images are recognized using the hybrid eight direction chain code features and 16-blocks based pixel count features. Basically, there are three steps of character recognition such as character segmentation, feature extraction and classification. In segmentation step, horizontal cropping method is used for line segmentation, vertically cropping method and bounding box is used for connected component character segmentation. In the classification step, the performance accuracy is measured by two ways such as KNN (K’s Nearest Neivour) classifier and feature based approach of template matching on 150 warning text signboard images.

Kyi Pyar Zaw | Zin Mar Kyu

Abstract

Character recognition is the process of converting a text image file into editable and searchable text file. Basically, there are three steps of character recognition such as character segmentation, feature extraction and classification. This paper mainly focus on the Myanmar character segmentation. Character segmentation is a vital area of research for optical character recognition. In this paper, the incoming text based images are segmented into lines, words and characters. Horizontal cropping is used for line segmentation and vertical cropping is used for vertically non-touching word and character segmentation. In a Myanmar compound word, there are one basic character and one or more extended characters. These basic character and extended characters may be connected or not according to the typing style or font style and Myanmar script nature. Therefore, it is difficult to segment these connected characters into individual characters. To solve this problem, we use block based pixel count and aspect ratio. This system can segment both touching characters and non-touching characters in text line image. Features are extracted from this segmenting characters. These individual characters are classified using eight directions chain code features and block based pixel count. Finally, the recognized text image is converted into editable text. In this paper, 92 characters in Myanmar script (34 consonants, 13 dependent vowels, 12 independent vowels, 1 punctuation mark, 10 digits, 8 medial, 5 compound medial, 3 tone characters and 6 compound words) are trained and different test line images that contain various words and characters are tested.

Kyi Pyar Zaw | Nu War

Abstract

This paper publicizes the touching character segmentation and recognition for the Myanmar warning text signboards. In the touching character segmentation step of this system, pixel connected component characters are extracted using connected component (CC) labeling with bounding box process. This system firstly search location zone of bounding box characters using minimum y-position and high features. This system develops a touching character segmentation technique that segment the Myanmar touching characters based on zone location. This segmentation technique uses the existing features such as number of holes, end points, horizontal black stroke count, vertical black stroke count, pixel count and the new features such as upper and lower sub-components in two horizontal zones, left and right sub-components in two vertical zones, end point existed zones. These features are further used in classification of segmented Myanmar characters. The system investigated on three types of datasets. First dataset includes 101 printed warning sign images. Second dataset includes 45 handwritten warning sign images that manually resized with various ranges based on visual font size, font style and number of text. The remaining dataset includes 152 real worlds Warning Sign Images (WSI) that automatically resized into 480×640 from 3120×4160 resolution and 640×480 form 4160×3120 resolution.