I proposed a mechanism addressing face anti-spoofing attacks based on DeepLabV3, achieving 93.7% accuracy in distinguishing real versus spoofed facial inputs. I explored scientific papers and conducted some research about existing segmentation models. The DeepLabV3 with a ResNet101 backbone was chosen. Also, creating suitable labels for the dataset, which originally contained textual labels was solved by developing segmentation-based labels for each image through a preprocessing stage. Interestingly, handling this research project in a remote group made me more self-motivated and independent than before.More importantly, the model remarkably could detect a human’s facial features such as eyes in spoofed images, highlighting the strong capabilities of semantic segmentation models to extract meaningful and informative patterns even in non-standard classification contexts. I am working on it to improve this framework, and I will share the code soon.

During my internship at Shenasa, I worked on computer vision projects related to face recognition, biometric systems, and object tracking . My work focused on studying different methods used to improve the reliability and efficiency of vision-based applications. I explored FAISS for large-scale face retrieval and analyzed how its indexing methods affect search performance when working with large datasets. I also investigated object tracking approaches, including SiamMask, DeepSORT, and ROLO, comparing their strengths and limitations in terms of accuracy and speed. In addition, I studied face quality assessment methods and biometric standards to understand how factors such as image quality, lighting conditions, and facial details influence recognition performance. This experience helped me gain practical knowledge of applying deep learning methods to real-world computer vision problems.