A lightweight coin detection and classification algorithm for low quality images

Coin detection and classification system is important in banks. When coin images are captured by low end line-scan camera with poor illumination, their quality will be relatively low. A lightweight algorithm is proposed for this kind of images. Firstly, a modified Canny operator is designed to filter vertical line segments to speed up the ellipse fitting steps. Secondly, a rotation invariant feature is applied to describe the feature of a coin. Initial experiments show that the classification accuracy of the proposed algorithm is about 92.6% for the data given. Introduction With the merits of durable and convenient carrying, coins are important circulating currency and play an important role in commodity transaction market. Therefore, coin detection and classification equipment is very important for banks. While various kinds of techniques have been adopted in modern coin detection and classification system, techniques based on images is a popular one. To improve the efficiency of coin detection and classification system, the conveyor belt carrying coins always runs at a relatively high speed, which means the system must capture coin images at a high frame rate. Matrix CCD/CMOS camera captures two dimensional images directly, but cannot capture images at high frame rate, especially for some low end cameras. On the contrary, line-scan digital camera can capture images with relatively high frame rate even with a low cost camera. Therefore, in the designed system, a low end line-scan digital camera is adopted to balance cost and speed. Different capture equipment produces images with different characteristics, which means different optimal detection and classification algorithm must be designed. Due to the low end line-scan camera adopted and poor illumination, the quality of the images is always relatively low. Figure 1 shows some sample data. (a) Normal coin (b) Game coin (c) Spoiled coin Fig.1. Low quality images sampled by the system In the images showed figure 1, the black strip indicates the conveyor belt. There are three kinds of objects to detect and classify. The first kind is normal coins as is shown in Figure 1(a), which means this kind of coins is genuine and can be put into circulation again. The second one represents counterfeit objects, including coins made for game machine, or some illegally produced fabricate coins. Some counterfeit objects have similar design as genuine coins, one of which is shown in figure 1(b). The third kind is the spoiled coins, which is distributed by the central government bank but is spoiled during the circulating stage. This kind of coins must be recycled by the bank to prevent it to be putted into circulation again. Therefore, the main task of a coin detection and classification system is to distinguish these three kinds of objects and signals the mechanical 2nd International Conference on Automation, Mechanical Control and Computational Engineering (AMCCE 2017) Copyright © 2017, the Authors. Published by Atlantis Press. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). Advances in Engineering Research, volume 118