Neural Image Thresholding Using SIFT: A Comparative Study

The task of image thresholding mainly classifies the image data into two regions, a necessary step in many image analysis and recognition applications. Different images, however, possess different characteristics making the thresholding by one single algorithm very difficult if not impossible. Hence, to optimally binarize a single image, one must usually try more than one threshold in order to obtain maximum segmentation accuracy. This approach could be very complex and time-consuming especially when a large number of images should be segmented in real time. Generally the challenge arises because any thresholding method may perform well for a certain image class but not for all images. In this paper, a supervised neural network is used to “dynamically” threshold images by learning the suitable threshold for each image type. The thresholds generated by the neural network can be used to binarize the images in two different ways. In the first approach, the scale-invariant feature transform (SIFT) method is used to assign a number of key points to the whole image. In the second approach,the SIFT is used to assign a number of key points within a rectangle around the region of interest. The results of each test are compared with the Otsu algorithm, active shape models (ASM), and level sets technique (LS). The neural network is trained using a set of features extracted from medical images randomly selected form a sample set and then tested using the remaining images. This process is repeated multiple times to verify the generalization ability of the network. The average of segmentation accuracy is calculated by comparing every segmented image with corresponding gold standard images.