H-C. LEE, T-Y. CHANG. Pill recognition using image processing techniques and neural
networks. Gerontechnology 2014;13(2):335; doi:10.4017/gt.2014.13.02.244.00 Purpose Medication
errors happen because of people’s mistake or misunderstanding. It is difficult for the general
public, especially the elderly, to recognize medicine simply by its appearance. In order to
solve this problem, many hospitals construct their own online drug recognition systems. Most
systems can only be searched with keywords containing the drug name or pharmaceutical
company, which is not a user-friendly system and may cause errors. In recent years, some
researchers have worked on the pill/drug recognition problem. The major features used for pill
recognition are the shape and color of the pill1
. In addition, imprint or more detailed information
about the pill is also used for pill recognition2-4
. The objective of this study is to use the widely
available web cam or smartphone for pill image acquisition and to facilitate pill recognition using
the proposed system with image processing techniques and neural networks. Method
We propose an automatic pill image recognition system that extracts features from a pill, such
as shape, color, and imprint. After multiple appearance and color features are extracted from
the image of a pill that is taken with a camera (Figures 1 & 2), these features will be processed
with the SimNet Neural Network for pattern classification. The proposed system will calculate
and show information for pills that have the top 5 closest similarity to the image. Results &
Discussion Experiments were conducted with over two thousand pill images, and promising
results were obtained. Multiple appearance features were extracted from a pill image that was
taken with a camera, and the features were processed by the proposed system with the SimNet
Neural Network for pattern classification. The proposed system showed the information for
similar pills that were correctly matched at ranks 1 to 5 (i.e., the 5 pills that were ranked most
similar to the image). We conducted experiments with the proposed approach on an image set
containing a total of 2,015 pill images. The success rate for the image of the target pill to be
ranked highest (i.e., most similar) is 92.66% and to be ranked in the top 5 (i.e., 5 most similar)
is 96.13%