Adjusting the thresholds to the recognised pattern in order to improve the separation between the recognised patterns

The aim of the following study was to examine the influence of image thresholding on the correctness of the pattern recognition in the grey scale images. The method based on moment invariants, which were the elements of feature vectors defining the features of the recognised object, was used by authors in order to recognise the objects. The paper presents the influence of image thresholding with histogram equalisation for exemplary images on the distribution of the distance between pattern vector and feature vector for every pixel of an image. The authors have paid a great attention to the fact that proper selection of the thresholds is significant for distinguishing given object classes. One could conclude from the results that adjusting the value of the thresholds to grey levels, which were in a searched object, could considerably improve object recognition. The proposed method was based on the analysis of the part of the image obtained from a camera. We assumed that the camera was mounted on the UAV and it watched the objects moving on the ground. This part of the image was selected in this way that it contained only a tracked object. We computed the histogram of this part of the image, equalised this histogram and computed new thresholds on the basis of the equalised histogram. Next step was to threshold the abovementioned part of an image and then to compute a pattern vector for this thresholded image of the object. In this way we obtained pattern vector which was used to recognize the object in the whole image. Since we wanted to use the pattern vector obtained on the basis of thresholded object image, we had to threshold the whole image (a terrain with moving objects) with the same thresholds. The examples presented in the paper showed that image pre-processing based on thresholding might improve the accuracy of the pattern recognition. It was achieved thanks to the thresholding which was conducted in a manner that guaranteed to distinguish features of the recognised object. In order to calculate the distance (ρ) between an object pattern vector and pattern vector obtained for a given point of the whole image, we used the Euclidean metric. We compared the values of the ρ obtained during recognition process which was conducted for both: the thresholded and original- not-thresholded image. The lower values of the ρ for the thresholded image meant that in this case the pattern vector described the features of the object better.