Thresholding techniques for multi-modal histogram into a sequence of Gaussian functions

Abstract

This paper describes thresholding for graphic images to reduce the information content for target detection. Firstly, a method has been described for partitioning a multi-modal histogram into a sequence of Gaussian functions. One single Gaussian function may not be adequate to approximate the whole dataset and hence requires creation of intervals to which a Gaussian equation may be fitted. The minimum RMSE value for Gaussian distribution fit has been used for creation first interval. Subsequent Gaussian fits are required to approximate the subsequent intervals in a sequential order. Thus a number of Gaussian functions are required to approximate the entire range. Points of inflection for each gaussian fit are considered as threshold point for that corresponding interval. A union of these thresholded images has been taken to get a single threshold image. Secondly, a method has been developed to extract target objects from the thresholded image. This method selects another threshold parameter by evaluating ratio of longest and the shortest dimensions for individual objects present in the binary image. It has been noticed that after applying these approach it significantly removes unwanted objects and detect the target. The developed method may be applied in surveillance where the objects histogram may vary because of the terrain (surrounding) and size may differ because of distance or shape of the vehicle. Experimental results demonstrate the robustness of the proposed method.