SEARCH RESULT

Retinoblastoma is a childhood cancer grows in retina. Although it could be treated in early stages, it can spread to nervous system and also other parts of the body and eventually may cause death in this situation. The prediction of survivability attracts a considerable interest and has been studied at different types of cancers, like breast, lung, colon and thyroid in literature by applying data mining methods. Data used in this study is obtained from The Surveillance, Epidemiology, and End Results (SEER) program which is an authorized data repository of cancer statistics. In our study, the survivability for retinoblastoma is predicted on SEER dataset consisting of 1258 patients by using data mining algorithms (support vector machines, logistic regression, multi-layer perceptron, naïve bayes, random forest and decision trees). Two strategies for imbalanced data which are over-sampling (synthetic minority over-sampling - SMOTE) and under-sampling are used. Results are analyzed and compared with the ones studied in other cancer types.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Gülistan Özdemir Özdoğan Hilal Kaya Baha Şen I. CANKAYA

Using Synthetic Aperture Radar images become popular in many military or civilian applications such as algorithm design, geo-referencing and Automatic Target Recognition. One of the main reason is SAR images can be obtained in any weather condition like rainy or cloudy weather even without daylight.However, Synthetic Aperture Radar (SAR) images contain multiplicative noise called speckle which makes analyzing images difficult. Therefore, there are many algorithms developed about despeckling SAR images in last decades. Each algorithm has strengths and weaknesses such as some algorithms work great in texture areas and some can work fine about homogeneous regions. To achieve more efficient result in despeckling SAR images, we proposed a method which uses 3 despeckling algorithms (SSD, MSAR_BM3D and FANS) and apply those algorithms in the regions which they are powerful. The proposed method splits a SAR image into smaller images and use Convolutional Neural Networks to categorize the sub images to find which algorithm is the best for that region. Afterwards, sub images despeckled using the algorithm which CNN selected and sub images come together and create the final despeckled image. The proposed method aimed despeckling of noises from the Synthetic Aperture Radar images more effective than the available despeckling algorithms.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Yusuf Şevki Günaydın Baha Şen

Handwriting character recognition has been a popular problem among scientists for a few decades. United States Postal Service can be given as an example for a company that uses the recognition of digits in real life environment consistently. USPS uses digit recognition system to extract digits from pay checks and fastens the process of sending and receiving checks. Handwriting character recognition problem can be divided into two categories. Online character recognition and offline character recognition. A recognition pattern mainly based on angle of the strokes of stylus is called online recognition. A system is called offline when system takes images as inputs and tries to predict characters from given images by applying machine learning methods. We have worked on offline character recognition problem in this project. Many machine learning methods have been proposed over the years for solving this problem. In this paper we implemented 6 most popular machine learning methods to solve offline handwriting character recognition problem and compare the performance results to decide which method gives best accuracy results under pre-defined conditions. We have selected 92255 images from NIST Special 19 Database and used them as input images during the training phase of the selected machine learning methods. These methods are SVM, Decision Tree, Bag of Trees, Artificial Neural Networks (ANN), Deep learning network with autoencoders and Convolutional Neural Networks (CNN). We implemented all of these methods and compare the performance of the results according to accuracy metric. The results obtained from the comparison is going to help in deciding which ML method should be used to solve Offline Handwriting Character Recognition problem.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Ş.G.KIVANÇ Ahmet Emin Baktır Baha Şen

Bu çalışmada Unity 3D oyun motoru ile araçların plaka takibinin simülasyonu yapılmıştır. Araçlara belirli bir yol güzergahı ve belirli bir hız verilerek kameralardan takibi yapılmıştır. Simülasyonda 16 kamera ve 3 araç kullanılmıştır. Hangi kameradan hangi aracın geçtiği ve araçların geçiş saati gibi bilgiler Microsoft SQL veritabanında saklanmıştır. Kameralardan elde edilen verilerin karşılaştırılması için iki aracın kamera ve saat bilgileri veritabanından çağrılmıştır. Bu iki kamera noktası arasındaki mesafenin tespiti için Dijkstra algoritmasından faydalanılmıştır. Araçların plakalarının aynı olması durumunda aralarındaki mesafe ve geçiş süresi bilgileri ile araçların bu mesafeyi katedip edemeyeceğinin tespiti yapılmıştır. Böylelikle araçlardan birinin sahte plaka kullanıp kullanmadığı belirlenmiştir.

Akademik Bilişim
AB

Salih Özkan Baha Şen Kemal Akyol