Improvement of facial biometric security requirements approach using partitional-based digital image clustering technique

Content Based Image Retrieval (CBIR) is a process to retrieve a stored image from a database by supplying an image as a query instead of text. It is used for, but not limited to applications such as facial recognition systems, medical diagnosis, architectural and engineering and information systems. Due to the benefits of CBIR engines, it is important to improve the efficiency of feature extraction techniques. Literally, biometrics security is concerned with identifying humans based on their unique physical and behavioural traits. In this research, CBIR approach has been used for Facial Biometric Security Image Processing (FBSIP). It extracts the biometric images to obtain three-octet 8-bits Red-Green-Blue (RGB) value where it used as medium to identify and recognise an image identity of user for criminal (wanted/suspect) profiling. The facial biometric image pixel value must be matched with the record stored in database or otherwise, the image identity is failed to identify. As per preliminary identified, the FBSIP requirements and specification is totally different from normal CBIR. Even though there are variation of Digital Image Clustering (DIC) technique developed, the DIC technique that focused on this research is only pixel based DIC technique. The DIC algorithm applied in experiment and analysis stage was determined from theoretical analysis stage. It is very certain that, analysis and discussion in that stage is adequate as a reason to nominate those DIC algorithms. On top of that, every DIC algorithm was designed and develop only for specific output. Hence, each DIC algorithm perform clustering employ own clustering computation technique. Other than that, some of DIC algorithm equipped with few parameters that would affect the result. The main objectives of this research are to study the existing CBIR techniques and algorithms in terms of their performance, to investigate the crucial FBSIP requirements for DIC algorithm patching, to nominate the compatibility a DIC algorithm with FBSIP requirements, and to validate the effectiveness of the proposed models during image retrieval for facial biometric security using RGB histogram for colour features extraction. During the experimental works, all DIC algorithms were tested with a certain number of biometric images (face). That experimental testing focused on image accuracy based on similarity index, accuracy rate percentage, and number of matched clusters. This novel model will be implemented in image similarity matching for update the criminal (wanted/suspect) profiling. In terms of the Object Segmentation Process, all objects are segmented based on distance from the cluster centre. From the theoretical analysis stage, the DIC algorithm used in the experiment and analysis stage was determined. It is certain that the study and discussion at that stage is sufficient to justify the nomination of those DIC algorithms. All DIC algorithms (K-Means, ISODATA, and K-Harmonic Means (KHM)) were tested with a dataset which is certain number of biometric images (human face) during the experimental process. The accuracy of a picture was tested using the Similarity Index (SI), accuracy rate percentage, and number of matched clusters. For the Facial Biometric Security Image Processing performance analysis, the Euclidean Distance or also refer SI has been compared. Based on that comparison, KHM produced higher query accuracy rate compared to the others which is 86.0%, ISODATA 80.5%, and the lowest rate is KMeans which is 79.5%. KHM not just better accuracy rate, it clocked fastest query process time, which is only 5 seconds, followed by K-Means, 7 seconds, and the slowest is ISODATA, 8 seconds. From the comparison result, it can be related that, objects that distributed into more matched clusters produced a better accuracy rate than fewer cluster distributions. For overall conclusion, it was found that the cluster distribution (number of k) is affecting the query time taken where objects distributed into more clusters is faster to perform query on it and also produced more accurate query result as the number of centroids is increase and place closer to objects. On top of that, the time ratio which is average time taken for each k is more efficient on more k set for DIC algorithm.