Browsing by Author "Amine, Chaima"

Now showing 1 - 4 of 4

Aerial forest smoke’s fire detection using enhanced YOLOv5
(Springer, 2023) Cherifi, Dalila; Bekkour, Belkacem; Benmalek, Assala; Bayou, Meroua; Mechti, Ines; Bekkouche, Abdelghani; Amine, Chaima; Halak, Ahmed
Forest fires around the world are the main cause of devastating millions of forest hectares, destroying several infrastructures and unfortunately causing many human casualties among both fire fighting crews and civilians that might be accidentally surrounded by the fire. The early detection of more than 58,950 forest fires and the real-time fire perception are two key factors that allow the firefighting crews to act accordingly in order to prevent the fire from achieving unmanageable proportions [1]. Forest fire detection is such a challenging problem for the current world. Traditional methodologies depend on a set of expensive hardware and sensors that might be not accurate due to some environment parameters and weather fluctuations. This paper proposes an accurate intelligent deep learning-based YOLOv5 model to detect forest fires from a given aerial images
Convolution neural network deployment for plant leaf diseases detection
(Springer, 2023) Cherifi, Dalila; Bayou, Meroua; Benmalek, Assala; Mechti, Ines; Bekkouche, Abdelghani; Bekkour, Belkacem; Amine, Chaima; Ahmed, Halak
The automated identification of plant diseases based on plant leaves is a huge breakthrough. Furthermore, early and accurate detection of plant diseases positively impacts crop productivity and quality. However, managing the accessibility of early plant disease detection is crucial. This work has environmental goals aiming to save plants from different threatening diseases by providing early detection of the affected leaves. We studied the performance of different Convolutional Neural Network (CNN) architectures in predicting 26 diseases for 14 plant species. The work studied the complexity of the system and compared the two main deep learning frameworks, TensorFlow and PyTorch, to get the most accurate results with higher accuracy. Using the “New PlantVillage Dataset” from Kaggle [1], the TensorFlow models achieved an accuracy of 90,94% for the basic CCN architecture, and 95,59% for the Transfer Learning architecture with VGG19. Whereas the PyTorch models achieved an accuracy of 93,47% for the basic CCN architecture, and 98,53% for the Transfer Learning architecture with ResNet34. Finally, after examining the feasibility of the model’s implementation and discussing the main problems that may be encountered, the models were deployed in a mobile application using the Tflite and torch mobile flutter SDK to let them as an internal feature in the mobile without the need for any access to the cloud, which is known as edge AI
Large language models (LLMs) exploration on pubMedQA and deployment
(University M’hamed Bougara : Institute of Electrical and Electronic Engineering (IGEE), 2025) Amine, Chaima; Cherifi, Dalila
Making LLM research accessible to everyone faces a big problem: fi ne-tuning these models requires so much computational power that most researchers and institutions simply can’t afford to participate in cutting-edge development. This creates a divide where only well-funded organizations can actually work with and improve LLM capabilities, while others are limited to pre-trained models that often miss the mark for their particular use cases. Optimization techniques like 4-bit quantization, LoRA, and aggressive frameworks like Unsloth claim they can solve this problem, but the thing is that current research only looks at how much computational power these methods save, not whether the optimized models actually work well in practice. This brings us to a critical question that nobody has really answered yet: when you optimize an LLM to run on limited hardware, which capabilities do you lose first, and how does this affect the model’s ability to handle domain-specific tasks like medical question answering? Our thesis systematically evaluate these trade-offs by looking at computational efficiency, task performance, and whether the models actually work in real-world scenarios. We used Llama 3.1 8B for medical question answering on the PubMedQA dataset, comparing base versus instruction-tuned models and testing recent advancement Zero-shot and few-shot learning as well as aggressive optimization against standard fi ne-tuning methods. To demonstrate the practical impact of our work, we deployed the selected model using a web interface, making it accessible for real-world testing and exploration. What we found was pretty clear: aggressive optimization techniques often make models computationally cheaper but practically useless. Unsloth’s optimization reduced computational requirements signifi cantly but made the fi ne-tuned model completely unreliable for medical question answering, while classical methods kept the model working well despite costing more resources. The idea that we can have both computational effi ciency and good model perfor- mance at the same time just isn’t true, and while optimization techniques are promising for researchers with limited resources, we still need balanced approaches that prioritize keeping themodel functional.
Study and implementation of u-net encoder-decoder neural network for brain tumors segmentation
(Springer, 2023) Cherifi, Dalila; Bekkouche, Abdelghani; Bayou, Meroua; Benmalek, Assala; Mechti, Ines; Bekkour, Belkacem; Amine, Chaima; Ahmed, Halak
Emerging advanced technologies have seen a revolution of applications into medical field, in all its aspects and sides, this has helped healthcare practitioners and empowered them in achieving accurate diagnosis and treatment, specifically with the evolution of computer Aided Diagnosis systems which use image processing techniques, Computer vision,and deep learning applied on different medical images in order to diagnose the image, or sections of the image with particular diseases or illnesses. Medical images of multiples organs or parts of the body (Liver, brain, kidney, skin, etc..) can today be visualized thanks to the advanced medical imaging techniques that exists in the market (MRI, CT, etc…) these technologies uses high energy in order to acquire high quality images but high energy can harm human cells, this is why we us low energy and with this used we get slightly low quality medical images, and here technology intervenes where we can use preprocessing techniques in order to increase image resolution prior to perform diagnosis either by doctor or CAD system. We present in this paper a computer aided diagnosis system that provides an automated brain tissue segmentation applied on 3D MRI images with its four different modalities (T1, T1C, T2, T2 weighted) of BRatS 2020 challenge dataset, by implementing a U-Net like deep neural network which provides information about classification of brain tissue into healthy tissue, Edema, Enhancing tumour, Non enhancing tumour. The model achieved an accuracy of 99.01% and dice coefficient of 47.95% after 35 epochs of training