Pusat Pengajian Kejuruteraaan Elektrik dan Elektronik - Tesis

Browse

Browsing Pusat Pengajian Kejuruteraaan Elektrik dan Elektronik - Tesis by Author "Ali Ahmed Abdullah, Al-Amuodi Ibrahim"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Publication
    Improving the efficiency of yolov7 tiny lightweight face detection model with ghost convolution and cbam attention mechanism.
    (2025-01)
    Ali Ahmed Abdullah, Al-Amuodi Ibrahim
    In recent years, significant advancements in face detection have been achieved through Convolutional Neural Networks (CNNs). However, a trade-off between model complexity and accuracy has been observed in existing approaches. Complex models offer high accuracy but are impractical for real-time or embedded applications, while lightweight models often suffer from low accuracy. To address this challenge, a lightweight face detector, GCB-YOLO (Ghost CBAM YOLO), is proposed to achieve an optimal balance between compactness and detection accuracy. The model incorporates the Ghost Module (GM) in its architecture to reduce size and the Convolutional Block Attention Module (CBAM) to enhance accuracy while maintaining a compact structure. The Sigmoid Linear Unit (SiLU) activation function is also integrated to boost performance. A user-friendly graphical interface has been developed to enable users to select different variants or modes for face detection. Significant size reduction is achieved through experimentation and optimization compared to established YOLO variants, with fewer parameters and computational resources required. Model size was reduced by 49.44% from 6.015M to 3.014M and speed in CPU increased from 4.265 FPS to 5.130 FPS, with a slight effect in accuracy as it dropped from 68.3% to 67.7% in the mAP50 and from 35.4% to 34.8% in the mAP50:95 Importantly, the mAP50 performance of existing models (except the original YOLOv7-tiny) is surpassed across various WIDER FACE sub-datasets, demonstrating competitive accuracy despite the streamlined architecture. While slightly lower mAP50:95 scores are observed compared to some competitors, a favourable trade-off between size and accuracy is exhibited, as evidenced by Multi Criteria Decision Analysis (MCDA). These findings suggest that the proposed GCB YOLO model is well-suited for deployment on resource-constrained devices, offering promising applications in CPU-based systems where efficient face detection with limited computational resources is critical.