Investigating the Impact of Data Parallelism and GPU Technology on Computer Gaming

Abstract:
According to the current design trends, multithreaded multicore processors will be ubiquitous in every device. In computer gaming, chip-makers are adding more cores to fulll the next generation performance requirements. A game engine has many `tasks' and data parallelism is an important technique for concurrent execution of these tasks. However, effective implementation of multithreaded computer games has challenges including concurrent/parallel processing, data and task level parallelism, and thread synchronization. In this paper, we investigate the impact of data parallelism and graphics processing unit (GPU) technology on multicore game engines. We implement a multi-object interactive game engine in an 8-core workstation using single-threaded model (STM) and various multithreaded models. We also implement a high quality DXT compression (a family of implementations of the S3 texture compression algorithm) using GPU technique. Experimental results show that multithreaded synchronous model with data parallelism (MSMDP) outperforms STM by reducing execution time up to 50%. Results also show that for 448-thread data parallelism, more than 81x speed up can be achieved by applying GPU computing.
Full Paper (in PDF)

---

Pectoral Muscle Elimination on Mammogram Using K-Means Clustering Approach

Abstract:
The performance of the Computer Aided Detection (CADe) system, to detect breast cancer, can be decreased due to some factors like presence of labels or other artifacts or pectoral muscle. Detection and segmentation of pectoral muscle can also help in image registration for further analysis of breast abnormalities such as bilateral asymmetry. In this paper, we proposed an algorithm based on K-means clustering to eliminate the triangular area of pectoral muscle. The reduction of irrelevant noise and unwanted artifacts are also performed using morphological preprocessing and seeded region growing (SRG) techniques. The method suggested for detection was tested over 322 images of 161 women taken from mini- MIAS database, out of which the proposed algorithm is able to eliminate pectoral muscle, showing 94.4% true positive value, from 291 images successfully. Results of pectoral muscle elimination are divided into three groups: Good (90.37%), Acceptable (8.07%) and Unexpected (1.5%).
Full Paper (in PDF)

---

Extraction of Potential Sunny Region for Background Subtraction under Sudden Illumination Changes

Abstract:
This paper proposes a novel background subtraction method robust for sudden illumination changes that often happen in outdoor scenes. The method first estimates regions where are sunny regions or would become sunny regions when the sun is not behind clouds, which we call "potential sunny regions". For the estimation, spatio-temporal analysis is applied to image sequences of the recent days of a target day considering the periodicity of the sun's movement. Once the potential sunny regions are obtained, they are used for judging if the sudden illumination change happens in the target scene. When it happens, then the illumination changes within the sunny regions are suppressed to obtain better subtraction results. Experimental results in several outdoor scenes show effectiveness of the proposed method.
Full Paper (in PDF)

---

Dynamic Blocks for Face Verification

Abstract:
Face recognition system is a computer based biometric information processing for automatically identifying or verifying a person from a digital image or a video frame. The significance of this research area is increasing day by day. Although the existing methods of face verification system perform well under certain conditions, there are still challenging problems due to several issues such as pose variation, facial expression variation, occlusion, imaging conditions, illuminations, size variations, age variations, orientations, etc. This paper addresses the problem of recognizing human faces despite the variations of pose and size. To handle these problems, we mainly focus on dynamic block size. Instead of uniform block, we propose Dynamic Size Blocks (DSB) considering most prominent face features such as eye, eyebrow, nose, mouth, chin, cheek, fore-head, etc., based on facial landmarks. In this feature based approach, we use a Dynamic Local Ternary Pattern (DLTP) for extracting facial feature information from each dynamic block. Then we perform a square-root of Chi-Square distance for similarity measurement of each block. We use a Support Vector Machine (SVM) classifier for face verification. We performed a comprehensive experimental evaluation on Labeled Faces in the Wild (LFW) dataset with restricted settings original images. Our proposed method has achieved an accuracy of 74.08% on all test images and 82.26% on dataset images excluding extreme pose variations.
Full Paper (in PDF)

---