Detection of the heterogeneities in the images using maximum bimagnitude estimation
Tutkimustuotos › › vertaisarvioitu
|Julkaisu||Telecommunications and Radio Engineering (English translation of Elektrosvyaz and Radiotekhnika)|
|DOI - pysyväislinkit|
|Tila||Julkaistu - 2020|
The novel technique developed for detection and recognition of the heterogeneities in the digital images is proposed and studied. The suggested approach is based on the evaluation of novel classification features in the form of the maximum bimagnitude value computed for the image pixel intensities contained in the local image segment. In order to examine the performance of the proposed technique test image contained the sharp contrast variations in the borders is designed. The mathematical model for computations by the Matlab software is developed for performance examination and the images contaminated by additive Gaussian noise with different variance values are studied. The receiver operating characteristic (ROC) parameter is used for evaluating the dependence of probability for true classified image areas on the false classified areas. Detection performance is assessed by the area under the curve (AUC) parameter. The proposed technique was compared with two common techniques based on the estimation of the local root mean squared (RMS) deviation values and quasi sweep values computed for the image pixel intensities contained within the slide window limits. It was demonstrated that bispectrum-based and local RMS deviation-based techniques provide very good detection of the image borders under the assumption of noise absence. However, at the same time, better performance was obtained by the proposed bispectrum-based technique in additive Gaussian noise environment and under increasing of noise variance value. It is also shown that the common technique based on the estimation of local mean squared values provides better performance under small noise variance values. Proposed and common techniques provide AUC values equal to 0.678 and 0.8468, respectively. The proposed technique provides a more efficient index under large noise variance values. Computer simulations results indicate that under noise variance value equal to 0.6, the proposed technique provides AUC value equal to 0.8748, i.e., proposed technique provides better performance as compared to common techniques.