Sistem pengecaman jantina amat berguna sebagai saringan bagi sistem pengawasan keselamatan berasaskan muka. Pelbagai sistem pengenalan jantina telah dicadangkan dan berjaya mencapai ketepatan yang tinggi bagi gambar muka yang ditangkap dalam suasana yang dikawal. Sistem ini juga menggunakan model yang kompleks dan besar. Oleh itu, projek ini akan membangunkan sistem pengenalan jantina bagi gambar yang ditangkap dalam suanasa yang tidak dikawal termasuk wajah orang yang memakai tudung, topi dan cermin mata dan pada masa yang sama memastikan sistem itu pantas dan bersaiz kecil. Sistem ini dibahagikan kepada pengesanan muka dan pengecaman jantina. Sistem ini menggunakan pengesan ‘haar cascade’ oleh MATLAB. Penambahbaikan dibuat menggunakan pra-pemprosessan, penyingkiran median dan membuat pengesanan pada imej yang diputar. Kajian menunjukkan peningkatan kejituan dan kadar pengesanan kepada masing-masing 92.54% dan 91.85%. Bagi pengecaman jantina, AlexNet, GoogLeNet, ResNet-18, dan VGGFace telah dilatih menggunakan pemindahan pembelajaran dan diuji pada set ujian CelebA. Hasilnya, model VGGFace memberikan ketepatan tertinggi pada 96.65% iaitu lebih tinggi daripada ketepatan 95% yang diperolehi oleh model yang sedia ada. Kemudian, model VGGFace dipangkas dan hasilnya kelajuan model meningkat hampir dua kali ganda, dan saiz model berkurang sebanyak 35.49%. Ketepatan model jatuh kepada 96.53%. Namun, Akhirnya, kedua-dua sistem digabungkan dan diuji dan ketepatan model VGGFace yang dipangkas jatuh hingga 95.432% yang menunjukkan bahawa ketepatan sistem ini dihadkan oleh sistem pengesanan muka.
_______________________________________________________________________________________________________
The gender recognition system is useful for various purposes such as a screening process for face recognition based security surveillance. Recently, a lot of gender recognition system had been proposed and managed to achieve high accuracy on constrained images. However, these systems are generally based on a complex and big model and can only perform well on a constrained condition. Thus, this project focus on developing a gender recognition system for an unconstrained image including faces of people wearing a hijab, a hat and sunglasses while making sure that the system is fast and small. The system is divided into face detection and gender recognition. The face detection part uses MATLAB’s haar cascade detector. Improvement is made by applying pre-processing, median rejection and making multiple detections on the rotated image. The result shows that the precision and recall improved up to 92.54% and 91.85% respectively. For gender recognition, pre-trained AlexNet, GoogLeNet, ResNet-18, and VGGFace are trained using transfer learning and are tested on the aligned CelebA test set. The result shows that VGGFace model gives the highest accuracy at 96.65% which is higher than 95% accuracy obtained by the existing model. VGGFace model’s convolutional layer are then pruned and the result shows that the model speed almost doubled, and the model size is reduced by 35.49%. However, the accuracy drops to 96.53% but considering the improvement made, the accuracy drop is insignificant. Finally, both systems are combined and tested on CelebA test set and the accuracy of the pruned VGGFace model drop to 95.43% which shows that our system accuracy is limited by the face detection system.
Computer vision based gender recognition using deep learning / Amirruddin Abdul Razak