Real-time in-plane rotation estimation of pharmaceutical tablets: a feature based approach

Visual quality inspection plays an important rule in quality control. With visual inspection it is ensured that the inspected surface meets the requirements; that the surface is free of irregularities, such as cracks, dents, and scratches. Detection of irregularities can, however, be a challenging task, especially on complex surfaces, where it is necessary to differentiate between surface irregularities and expected surface variations. Differentiation is done by comparing a sample with a reference model, which requires an accurate spatial alignment between the two. In this paper we focus on visual quality inspection of pharmaceutical tablets. While inspected, tablets are mechanically constrained, thus alignment simplifies to in-plane rotation estimation. For in-plane rotation estimation we propose a method using histograms of oriented gradients (HOG) and a nearestneighbor regression in HOG feature space. Method was evaluated on four datasets of pharmaceutical tablets, varying in size, shape and color. The results show that the proposed method is superior in robustness, with comparable accuracy to the methods previously used for rotation estimation of pharmaceutical tablets.