IR-EAR: A Dataset of Left and Right Ear Images: Case Study of Iranian Subjects/ Persian Race
Description
To the best of our knowledge, the IR-EAR dataset is the first Iranian ear dataset with high-resolution images, a substantial number of Iranian individuals, and detailed demographic and biometric attributes. We collected this dataset at Azad University in Mashhad, Iran. Most participants were bachelor's and graduate students, and we also collected some images from staff. The dataset includes left and right ear images of 406 Iranian individuals: 201 males and 205 females. We built a custom box to create a consistent environment for capturing ear images and to facilitate collection, saving, and cropping (a photo of the box is provided in the data description file). The distance between the participant's ear and the camera was approximately 13 cm. For each participant, we captured direct photographs of both left and right ears using an iPhone 13 Pro Max. We took images with an ultra-wide camera equipped with a 13 mm f/1.8 lens at a resolution of 2268 × 4032 pixels, saved as PNGs, and approximately 3.1 MB each. After image acquisition, we reviewed the images and selected the highest-quality image from available duplicates for each ear. We excluded incomplete, blurred, duplicated, poorly illuminated, or otherwise unsuitable images. We then performed preprocessing, including quality assessment, background removal, cropping, and labeling to improve consistency and suitability. We used a systematic labeling procedure for consistency. Each final image label consisted of three components: a sample identifier, an ear-side indicator, and the unique numerical identifier generated by the smartphone. For instance, an image originally named IMG_2173.HEIC was renamed S001_R_2173. Here, S denotes the sample, 001 the three-digit participant identifier, R the right ear, and 2173 the unique numerical identifier generated by the smartphone. The letter L denotes left-ear images. The dataset features are organized in a CSV file containing demographic and biometric attributes for every participant, including the identifiers of the left and right ear images, gender, age, height, weight, marital status, ear-related medical conditions, racial background, ambidexterity, eye color, occupation, and hearing aid usage level. The final dataset is divided into 4 ZIP files for easier access. The IR-EAR dataset supports research on computer vision and machine learning for the Iranian population and provides a resource for ear-based biometric analysis, forensic investigations, and training machine learning models for Iranian populations and populations with similar craniofacial characteristics, such as individuals from Central Europe and parts of the Middle East. In multimodal biometric systems, ear biometrics can also be combined with modalities such as facial or iris biometrics to improve accuracy. The source code used for the validation and analysis of this dataset is publicly available on GitHub: [https://github.com/Dr-kamel/GenderClassification-EarShape/].
Files
Institutions
- Islamic Azad University, MashhadRazavi Khorasan, Mashhad
- Islamic Azad University, TehranTehran, Tehran
- Technical and Vocational UniversityTehran, Tehran