Invariant Image-Based Currency Denomination Recognition Using Local Entropy and Range Filters

• Published in: Entropy (Basel, Switzerland) Vol. 21; no. 11; p. 1085
• Main Authors: Anwar, Hafeez, Ullah, Farman, Iqbal, Asif, Ul Hasnain, Anees, Ur Rehman, Ata, Bell, Peter, Kwak, Daehan
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01-11-2019; MDPI
• Subjects: Algorithms; Aspect ratio; Currencies; Entropy; image entropy; Image filters; image processing; image segmentation; Invariants; Machine learning; Methods; Neural networks; Object recognition; Principal components analysis; Support vector machines; Visual impairment
• ISSN: 1099-4300; 1099-4300
• DOI: 10.3390/e21111085

We perform image-based denomination recognition of the Pakistani currency notes. There are a total of seven different denominations in the current series of Pakistani notes. Apart from color and texture, these notes differ from one another mainly due to their aspect ratios. Our aim is to exploit this single feature to attain an image-based recognition that is invariant to the most common image variations found in currency notes images. Among others, the most notable image variations are caused by the difference in positions and in-plane orientations of the currency notes in images. While most of the proposed methods for currency denomination recognition only focus on attaining higher recognition rates, our aim is more complex, i.e., attaining a high recognition rate in the presence of image variations. Since, the aspect ratio of a currency note is invariant to such differences, an image-based recognition of currency notes based on aspect ratio is more likely to be translation- and rotation-invariant. Therefore, we adapt a two step procedure that first extracts a currency note from the homogeneous image background via local entropy and range filters. Then, the aspect ratio of the extracted currency note is calculated to determine its denomination. To validate our proposed method, we gathered a new dataset with the largest and most diverse collection of Pakistani currency notes, where each image contains either a single or multiple notes at arbitrary positions and orientations. We attain an overall average recognition rate of 99% which is very encouraging for our method, which relies on a single feature and is suited for real-time applications. Consequently, the method may be extended to other international and historical currencies, which makes it suitable for business and digital humanities applications.