Classification of wrinkles on the forehead and around eyes

People communicate through their faces. They express themselves, feelings, give clues about their gender, age etc. via their faces. Face carries valuable information. Therefore it is important to be fulfilled about their appearance for people. It helps people to increase pleasure that they gain from life. Facial wrinkles appear in time due to loss of elasticity. On the other hand there are environmental issues that cause wrinkle on the face such as ultra-violet exposure or facial expressions (yawning, smiling etc.). The wrinkles can have different severities. While some of them are barely visible, some of them can be extremely deep. According to the clinicians (such as dermatologists, plastic surgeons), the treatment of the wrinkles is really based on their severity. The treatment for the mild wrinkles may not help to remove extreme wrinkles. Therefore, the classification of the wrinkle is a crucial point to find the optimal treatment. In addition, one cannot classify the wrinkles, if the wrinkles are not detected and localized. For this purpose, cosmetic researchers have performed various studies. Their main concern is to provide standard systems to classify wrinkle severity. A group of dermatologists, plastic or aesthetic surgeons come together to grade wrinkle severities of the collected images. The grading process happens in two sessions mostly. The inter-rater reliability is calculated after the first session. The raters that have highest reliabilities are selected for the second session. Then they grade the wrinkles again. In the end the wrinkle severity systems are obtained. This process has two main drawbacks. First of all the grading systems are strongly related to the raters. If the raters are changed, the scores can be affected. In addition, the system is not automatic to reproduce the results in anywhere at any time. When the aim is to develop reliable, repeatable and standard systems, there are significant studies in the computer society. However, the research is about the detection of the wrinkles mostly. Classification of the wrinkles is not a well studied area. The annotated wrinkle locations are being tried to find by the automatic systems. The studies are typically for the forehead wrinkles mostly. By annotating the collected images, Gabor filtering, Hessian filtering, Hybrid Hessian filtering etc. methods are being applied to detect wrinkles. The aim of this study is to build a standard, repeatable, reliable, automatic system to detect and classify wrinkle severities. A database is required which has samples and labels for the wrinkles. Firstly, a database was collected for this purpose. The images were taken from three different parts of the face. These regions are forehead, left and right eye. The images were taken from the same view under the same light. The layout of the camera was used to provide the standardization. The images were taken from the volunteers. The volunteers were not wearing any make up and they also did not have surgery in the last three years. The images were taken from 29 subjects with and without flash light. The forehead images were labeled according to the eye brows. The eye images were labeled using eye and the ear points. These points were used to align and crop images for grading. The last version of the images only includes the wrinkle parts of the face to exclude any bias. The prepared images were graded by five female and five male volunteers. The volunteers were used their point of view while scoring. The categories that will be used were specified by using the previous research. Hence, the specified number for the categories was three. The absence of the wrinkle and the mild wrinkles belong to the first category while the moderate wrinkles belong to the second category. The extreme and deep wrinkles belong to the third category. Before scoring, raters took a look to the whole images to have an initial information about the wrinkles. Then, they were asked to grade the wrinkle severities according to their observations. To calculate inter-rater reliability the Kappa (κ) coefficient were used. According to the calculated κ coefficients, the reliabilities were found 0.63, 0.79 and 0.69 for the forehead, left and right eye respectively. The results were found substantial to use. By using the scores, the images were divided into train and test sets. Three-fold were defined to use for the reproducibility. The train sets included the same number of the samples. The person independency was also taken into account. In this study two different feature extraction and classification methods were used to detect and classify the wrinkles. The first method uses Local Binary Pattern (LBP) to extract features and Support Vector Machine (SVM) to train a model. 16 × 16 pixels and 8×8 pixels patches were used for Local Binary Pattern. The uniform LBP features were also used because of their informativeness. The features were trained using linear SVM and RBF SVM. The second feature extraction method was Gabor filter. However, the images were smoothed by Gaussian filter. Then the Gabor filter was applied to have its phase value. The erosion, dilation and Sobel edge detection were applied to the phase value in order. Finally, Connected Component Analysis were used to find the count of the wrinkles on the images to use as feature. The features were classified by using K Nearest Neighbour algorithm with majority voting. When the experiments were examined, Gabor features outperformed LBP features in every category. While the highest accuracy of the classification of the wrinkles on the forehead were 64.5%, the accuracy of 79.3% were obtained using Gabor features. The accuracy was 64.4% for the classification of the wrinkles on the left eye when LBP was used. However the accuracy of 73% was obtained using Gabor features for the same region. The accuracies of the classification of the wrinkles on the right eye were 63% and 78.9% for the LBP and Gabor features respectively. While the LBP features mixed the second and third category with each other mostly, Gabor features were more successful to classify first and third categories. Gabor features mixed second category with the third category rarely. The wrinkles around the mouth were also shot while collecting the database. However the classification results were not good. The images should have taken closer and with a proper angle. Then the wrinkle length was added to the wrinkle number for the classification. Nonetheless, these two features could not provide sufficient information for accurate classification.