A new imprinted tablet recognition algorithm using polar transform and neural networks

The paper presents a new imprinted tablet recognition algorithm using polar transform and neural networks. The purpose of this algorithm is to determine if inspected tablets have the same imprinted symbol as a reference tablet or not. The algorithm consists of two phases namely neural network training phase and imprinted tablet recognition phase. In the neural network training phase, firstly blister images are captured by a CCD camera. The blister image is partitioned into separate tablet images. A set of sample tablet images with the same imprinted symbols are chosen. Imprinted symbol in the sample tablet is filtered out using image processing operations. The sample tablets are then rotated to be aligned with a reference tablet using polar transform. Next imprinted symbol features which are the total number of white pixels in a rectangle window are extracted to train neural networks. In imprinted tablet recognition phase, the inspected tablet image is first rotated to be aligned with the reference image, then features of the image are extracted and fed to the inputs of the trained neural networks. The output of the trained neural network allows to determine if the inspected table has the same imprinted symbol as the reference tablet or not. Experimental result shows that the proposed algorithm can recognize imprinted tablet accurately, the result is robust to uncertainties such as tablet rotation and lighting condition.