Feature based face detection for unconstrained images

• मुख्य लेखक: Tioh, Keat Soon
• स्वरूप: थीसिस
• भाषा: अंग्रेज़ी; अंग्रेज़ी
• प्रकाशित: 2021
• विषय: T Technology (General); TA Engineering (General). Civil engineering (General)
• ऑनलाइन पहुंच: http://eprints.utem.edu.my/id/eprint/26025/1/Feature%20based%20face%20detection%20for%20unconstrained%20images.pdf; http://eprints.utem.edu.my/id/eprint/26025/2/Feature%20based%20face%20detection%20for%20unconstrained%20images.pdf

Face detection for unconstrained images often encounter issues like background variation, pose variation, facial expression, occlusion and noise. Face detection utilises two main methods; feature based and image based methods. The feature based method benefits from rotation independence, scale independence and quick execution time as compared to the image based method. Feature based method utilises skin colour, facial and blob features. Current research on feature based method often emphasises on Viola Jones (V-J) face detection and is only limited to the in-plane rotation of positive or negative forty-five degrees. However, the utilization of V-J face detection with the inclusion of noise is a challenge because the image of other objects will often be mistaken for faces thus resulting in false detections. This thesis focuses on pose variation and noise challenges of unconstrained images and will cover three techniques for V-J face detection for unconstrained images, namely the combination of V-J face detection with rotation enhancements, Bicubic interpolation and ratio Scale Invariant Feature Transform (SIFT). In this thesis, these three techniques play different roles in face detection. The first technique begins with the rotation of the image file at thirty degree steps until it reaches a total rotation of three hundred and sixty degrees. At each thirty degree step, V-J face detection is applied, which in turn covers more angles of a rotated face. The second technique, Bicubic interpolation, corrects distorted images. The third technique, ratio SIFT, is a proposed post-processing to eliminate false detection for unconstrained images. Robust feature detection in scaling and invariant rotation is utilised in the above techniques to aid in the detecting of faces in images. Different face detections have been recommended for the unconstrained grey images and unconstrained colour images respectively with in-plane rotations and some with multiple faces. The images utilised for testing and evaluation in this thesis originated from Carnegie Mellon University (CMU) unconstrained grey images with in-plane rotations and Face Detection Data Set and Benchmark (FDDB) unconstrained colour images with multiple faces datasets. Fifty CMU datasets with twelve rotations on each image and various permutations resulted in six hundred test pattern images have been performed. Furthermore, another six hundred test pattern images from FDDB were also evaluated. These images have been measured through correct detection rate, true positive and false positive. The results from these measurements indicate that the proposed feature based face detection technique, focused on the V-J face detection method, for unconstrained images has the ability to detect rotated faces with high detection accuracy which in turn reduces false detections. In conclusion, the proposed enhancements will improve the current V-J face detection technique and overcome future challenges for unconstrained images.