اميرمحمد خادم حسيني

In recent years, semantic segmentation has emerged as one of the fundamental problems in computer vision, gaining particular importance in real-time applications such as autonomous driving. The main challenge in this field is achieving high accuracy while operating under computational an‎d hardware constraints. In this research, a lightweight an‎d efficient architecture based on the LMIINet network was trained using quantization-aware training (QAT) with the QKeras library, an‎d then implemented on FPGA hardware using the CGRA4ML framework. The Cityscapes dataset was selec‎ted as the reference for training an‎d eva‎luation. During this work, the necessary modifications were applied to adapt the model to CGRA4ML constraints, including simplifying skip connections, employing hardware-friendly operations (such as depthwise-separable an‎d 1×1 convolutions), an‎d redesigning parts of the Flatten Transformer. Furthermore, all layers were quantized to 8-bit precision, which not only reduced the memory footprint by a factor of four but also enabled efficient fixed-point computations on FPGA. The results demonstrated that with approximately 90% pixel accuracy an‎d 45% mIoU, the model can operate in real-time at 20 frames per second (50.104 ms latency) on the ZCU104 board. This research shows that CGRA4ML, with its flexibility in mapping modern layers an‎d its use of off-chip memory for skip connections, is a practical option for implementing advanced semantic segmentation networks an‎d other intelligent models on FPGA in real-time applications. However, additional modifications are required to further adapt semantic segmentation models for CGRA4ML deployment in order to optimize performance.

تقسيم‌بندي معنايي , آرايه‌ي دروازه‌ي ميداني برنامه‌پذير , كوانتيزاسيون

Semantic Segmentation , FPGA , Quantization

Amir Mohammad Khadem Hoseini

Dr Sattar Mirza Kochaki