Islem Touati

Our project aimed to tackle the growing challenge of detecting tampering in images and videos, a crucial issue in fields of criminalistics. We structured our approach around three key components: image processing, statistical analysis of image characteristics, and deep learning-based methods. Key Contributions: 1. We applied basic visual analysis techniques, enabling investigators to perform a thorough visual inspection before advancing to more sophisticated methods. 2. Statistical Analysis of Image Properties: We developed statistical methods to analyze image characteristics, identifying inconsistencies in the image formation process. This passive-blind approach was critical in detecting subtle tampering without needing prior information about the image, providing a flexible solution across various scenarios. 3. Deep learning-based methods: We implemented the CRISP-DM framework, selecting benchmark datasets, performing Exploratory Data Analysis (EDA), and ensuring data quality for robust model performance. For modeling, we built a classification pipeline using a bagging approach, combining Five models, including EfficientNet and InceptionNet, to achieve greater reliability and adaptability. For image segmentation, we enhanced a UNet model with a VGG encoder, improving the localization of tampered areas. The solution was successfully deployed as a Streamlit application, making it easily accessible for both technical and non-technical users. This interactive platform enables users to upload images and receive real-time tampering analysis, bringing our research closer to practical real-world applications. Moving forward, we plan to extend our system to include video tampering detection, allowing us to tackle more complex media formats and broaden the system's impact.

Within the company BK Fire, I worked on the detection of PCB defects, specifically open circuits and short circuits. The primary objective was to design an effective solution to accurately identify and classify these two types of defects. The main steps of this project included: Creation and acquisition of datasets, necessary for training the detection models. Annotation of data, carried out with tools such as LabelImg, CVAT (online), and the auto-annotation tool Autodistill. Application of Data Augmentation techniques (adjusting brightness, contrast, injecting noise, and geometric transformations such as zoom and rotation, etc.) to enhance data diversity. Detection of defects using the YOLOv8 model, combined with transfer learning techniques, to improve performance. This project enabled the optimization of PCB quality inspections by reducing errors and improving the reliability of quality control. Automation accelerated the identification of defects while paving the way for new detection methods.

During my internship, I worked on the development of a solution for the automatic classification of Algerian legal articles and texts, categorizing them into three classes. This project used the SetFit method (few-shot learning) with the CamemBERT-large model, pre-trained for French. My main contributions were as follows: Designing a system for the automatic classification of Algerian legislative texts, using transfer learning techniques and libraries such as Hugging Face and TensorFlow. The objective was to classify the articles as modified, completed or repealed, while improving the accuracy of this detection. Additionally, I developed a solution to link these articles (modified, completed or repealed) to their original versions, using the Regex library and transfer learning models, to facilitate the consultation and traceability of legislative modifications. The project was successfully deployed, enabling the automation of the legal modification detection process. This allowed legal professionals to save time and improve the accuracy of their research, thus enhancing the efficiency and reliability of legal research.

I developed a machine learning system to predict the existence, volume, and associated costs of claims for each contract. This solution was deployed using Streamlit and is designed to provide actionable insights into potential contract risks, enabling more informed decision-making. The system helps streamline risk assessment and claims management processes, ultimately reducing uncertainty and improving overall efficiency.

- Discover the working conditions of industry. - Understand the activities of the company and how it works.