Varin Chouvatut

Virtual piano with real-time interaction using automatic marker detection

This paper aims to apply the object detection used in Augmented Reality (AR) technology to a real-time application for musical instrument, a virtual piano. The proposed application does provide the sounds of musical notes and also display the corresponding notations of the playing notes. Both features allow people with hearing disability or muscular weakness or even ones who cannot exert their pressure on the general keyboard instruments to play music, due to their physical disabilities. The experimental results demonstrated that the error of the playback sounds is only 0.5 percent which is only 2 times measured from tapping 50 markers for 400 times in real time. However, all errors are because of the wrong positions of the players hand as some of his fingers accidentally covered over the unexpected markers during playing the virtual piano.

Training set size reduction in large dataset problems

Classifiers have known to be used in various fields of applications. However, the main problem usually found recently is about applying a classifier to large datasets. Thus, the process of reducing size of the training set becomes necessary especially to accelerate the processing time of the classifier. Concerning the problem, this paper proposes a new method which can reduce size of the training set in a large dataset. Our proposed method is improved from a famous graph-based algorithm named Optimum-Path Forest (OPF). Our principal concept of reducing the training sets size is to utilize the Segmented Least Square Algorithm (SLSA) in estimating the trees shape. From the experimental results, our proposed method could reduce size of the training set by about 7 to 21 percent comparing with the original OPF algorithm while the classifications accuracy decreased insignificantly by only about 0.2 to 0.5 percent. In addition, for some datasets, our method provided even as same degree of accuracy as of the original OPF algorithm.

Under-sampling by algorithm with performance guaranteed for class-imbalance problem

Class-imbalance problem is the problem that the number, or data, in the majority class is much more than in the minority class. Traditional classifiers cannot sort out this problem because they focus on the data in the majority class than on the data in the minority class, and then they predict some upcoming data as the data in the majority class. Under-sampling is an efficient way to handle this problem because this method selects the representatives of the data in the majority class. For this reason, under-sampling occupies shorter training period than over-sampling. The only problem with the under-sampling method is that a representative selection, in all probability, throws away important information in a majority class. To overcome this problem, we propose a cluster-based under-sampling method. We use a clustering algorithm that is performance guaranteed, named k-centers algorithm, which clusters the data in the majority class and selects a number of representative data in many proportions, and then combines them with all the data in the minority class as a training set. In this paper, we compare our approach with k-means on five datasets from UCI with two classifiers: 5-nearest neighbors and c4.5 decision tree. The performance is measured by Precision, Recall, F-measure, and Accuracy. The experimental results show that our approach has higher measurements than the k-means approach, except Precision where both the approaches have the same rate.

Efficiency comparisons between k-centers and k-means algorithms

This paper proposes an under-sampling method with an algorithm which guarantees the sampling quality called k-centers algorithm. Then, the efficiency of the sampling using under-sampling method with k-means algorithm is compared with the proposed method. For the comparison purpose, four datasets obtained from UCI database were selected and the RIPPER classifier was used. From the experimental results, our under-sampling method with k-centers algorithm provided the Accuracy, Recall, and F-measure values higher than that obtained from the under-sampling with k-means algorithm in every dataset we used. The Precision value from our k-centers algorithm might be lower in some datasets, however, its average value computed out of all datasets is still higher than using the under-sampling method with k-means algorithm. Moreover, the experimental results showed that our under-sampling method with k-centers algorithm also decreases the Accuracy value obtained from the original data less than that using the under-sampling with k-means algorithm.

Detection of Brain Tumor from MR Image Sequence Using Image Segmentation and Blob’s Centroid

This paper proposes a method to search for a probable tumor in magnetic resonance (MR) images of a human brain. Typically, a tumor can be found in some contiguous images of the MR sequence and positions of its appearance in such contiguous images usually have similar centroid thus their corresponding projections should be able to be detected automatically in order to support a user or a doctor for further diagnosis. Once region of a probable tumor is detected, matched checking between a pair of contiguous MR images can be done and relabeled to indicate the same area of the tumor amongst sequential images. Any regions without match between contiguous images are initially considered as irrelevant components and will not be analyzed further unless the doctor indicates otherwise. Then, ratio of tumor to brain is calculated to support as an initial diagnosis of tumors appeared in an MR image sequence .

Graphical representation of the whole sequentially MRI images in a single view image sequences of human's whole head

Typically, a sequence of the Magnetic Resonance Imaging (MRI) images is composed of a certain number of images projecting some internal organs of a human, such as brain and eyeballs of the humans head, which is the case chosen for our demonstration. Each of MRI images in the sequence presents only a thin layer of the whole head. The image processing techniques proposed in this paper aims to allow all such sequential images to be visible through a single view. In other words, the whole head of a human can be visible in just one image and thus looks as a three-dimensional view of the head. Unfortunately, there may be some deviation in positions even between contiguous images in the sequence. Centroid of the humans head appeared in each image should be measured. To ensure a centroids position is estimated well enough so that centroids of all sequential images are not so much deviated from each other, searching for the centroid of a humans whole head is done based on an approximate convex shape rather than a circular shape as usual. From our experimental results, there is no significant deviation of centroids between contiguous frames as expected.

Brain tumor's approximate correspondence and area with interior holes filled

Measuring area of tumor in humans brain from only single image may provide incorrect information for further diagnosis. Generally, a doctor or an expert must examine a brain tumor from several sequential MRI images to conclude its size or the severity level of patients illness. To imitate the way a doctor diagnosing such case in a real situation, some digital image processing techniques are proposed and applied in order to provide support for a tentative or an initial analysis to the doctor. Thus, correspondence of appearances of a tumor presented in all MRI images should be linked and considered. In image processing, a closed area can be seen as an object and based on the similarity of its interior shadings, the objects centroid can be estimated. Unfortunately, although an objects centroid may be calculated even there exists slightly different shadings which are still considered as having similarity inside the closed shape of the object, only a small hole can cause deviation of computed centroid from its expected position. Since the typical thresholding techniques still leave a hole whose area has a certain amount of different shading from the major shading of the objects area. Thus, we proposed a number of image processing techniques for the purpose of tumor area approximation. Moreover, the proposed methods include a correspondence technique would also support multiple-object detection and linking centroids of the same object, which is a brain tumor in this case, presented in a pair of contiguous images.

A search for geometric-shape objects in a vector image: Scalable Vector Graphics (SVG) file format

This paper proposes a model for detecting objects with geometric shapes in an image whose file format is SVG (Scalable Vector Graphics). The SVG is known as a type of vector images. A major interesting feature of SVG is that components in victor images are created using mathematical theories and equations, in addition, vector images have many more advantages over raster images in various aspects. Moreover, an objects details in a humans sight, a triangle for example, is looked as it has three corners. But the true shape created in SVG of the object may be different from the humans limited sight. Observing these gaps or limitation, we thus propose alternatively simple methods for processing on objects with geometric shapes directly in a vector image without loss of any details in its original form.

Pattern extraction from northern Thai fabrics using flexibly matching segments: Sarong Teenjok and Lanna textiles

This paper proposes a method with flexibility in detecting and recognizing northern Thai fabrics which usually include various delicate patterns or traceries. Due to delicacies in unique patterns of northern Thai fabrics, especially ones called Sarong Teenjok or even other Lanna textiles, flexibility in searching and detecting such patterns is required. Authors thus propose a method using some parameters allowed to be adjusted in order to add flexibility in pattern recognition. Thus, some small deviation, orientation, or different shadows and shades of the appeared traceries does not cause much problem. From experimental results, the proposed method still works very well in various challenging situations as mentioned, for example, in cases of having shadows in the input image, natural skewness or unnatural orientation of the fabrics, or some small deviation in patterns colors of the traceries.

Multi-view hand detection applying viola-jones framework using SAMME AdaBoost

Human hand detection is one of a popular researches in the field of object detection. One obvious problem of hand detection is about orientation angles of the hand position. That is, most detectors cannot detect a human hand lying in various orientation angles recently. Detecting hand with various orientation angles can be done using decision tree as a degree estimator. Using the decision tree as a degree estimator can cause the over-fit problem. In this paper, we propose the use of SAMME algorithm instead of the decision tree to prevent the problem. Moreover, from our experimental results, using SAMME as the degree estimator provides detection rate not less than using decision tree as the degree estimator. The results obtained from using SAMME algorithm as the degree estimator show that our detection rates increase by 4.01% (from 78.71 to 82.72) and 8.75% (from 77.82 to 86.57) on two experimental datasets. Their false positive rates decrease from 1 out of 2,959 to 1 out of 3,805 in the first dataset and from 1 out of 2,663 to 1 out of 4,566 in the second dataset, both of which are very low.