(Dept. of CS) Hitesh Adhao (M.Sc, 2nd Year) Abstract To quantify similarity in different images based on time-series analysis. Array to time-series conversion Local maxima detection-joining
Calculation of cumulative angle
Introduction Why? Pattern recognition is continuously evolving with the advancements in image analysis and machine vision. Human vision and perception can be considered as the minimum benchmark for all techniques used in machine vision. Under specific conditions, the human perception cannot be matched equally by a traditional mathematical parameter. This necessitates to explore new mathematical formulations. Is there any method for Pattern Recognition? Euclidean Absolute difference distance Correlation All above methods would yield relatively poor similarity due to change or shift in pixels position inside the image. 2D correlation coefficient for first two patterns would yield a mere 5% similarity. Human perception, on the other hand, probably identifies the second image as a slightly distorted version of the first image. Proposed Methodology Proposed Methodology Both patterns are the same, angles (EAB, EBA, ECD, and EDC) are correspondingly equal in both patterns, irrespective of their orientation. The core idea of the proposed method is to examine the relationship between successive pixels in an image. Figure 3 shows the proposed methodology starting from reading images to parameter calculation. Cumulative Angle (CA) Cumulative Angle (CA) on the other hand consists of three values: Two CA's for corresponding images; and their absolute difference. The absolute difference is a direct pointer to dissimilarity in the patterns. A higher difference means a greater mismatch. Data transformation The conversion process starts with scanning each row from top left corner to lower right side of the image. This generates a (1 X MN) array from an (MxN) gray scale image. After computing local maxima, a corresponding vector containing only maxima points is formed. This new vector represents a curve joining all the local maxima. A cumulative angle is calculated by first considering two successive maxima points, and then finding out the angle corresponding to slope of line joining these two points. Angle formation for lines AB and BC respectively are shown, where A,B,C are maxima in succession. The decision to choose local maxima points instead of pixels themselves as source of vector quantity (angle) is to reduce the computational complexity of the method, since there will be less number of angle formations. For example, there are 8 variations from index 1 to 10 for original time series, in comparison to 4 variations when local maxima are considered. A cumulative angle is calculated by first considering two successive maxima points, and then finding out the angle corresponding to slope of line joining these two points. Angle formation for lines AB and BC respectively are shown, where A,B,C are maxima in succession. The decision to choose local maxima points instead of pixels themselves as source of vector quantity (angle) is to reduce the computational complexity of the method, since there will be less number of angle formations. For example, there are 8 variations from index 1 to 10 for original time series, in comparison to 4 variations when local maxima are considered. Results and Discussion The parameters included for comparison are: Correlation coefficient Euclidean distance Cumulative Angle (CA) Conclusions The approach enables more human like judgment in case of specific pattern matching. The comparison method can be conceptually applied and verified for two different sized images. The process time depends on image size making it more favourable for small and medium sized image/pattern comparison. The proposed technique possess great potential in the areas of automatic archaeological literature analysis, robotics vision etc. References Paawan Sharma ,University of Petroleum & Energy Studies, Dehradun, Uttarakhand 248001, India Mukul Kumar Gupta UPES, Dehradun, 248001, India Amit Kumar Mondal UPES, Dehradun, 248001, India Vivek Kaundal UPES, Dehradun, 248001, India