Real-time edge face detection using a lightweight multi-task cascaded convolutional neural network with hardware-software co-design

This study presents a hardware-software co-design framework for real-time face detection on resource-constrained edge platforms using a lightweight Multi-Task Cascaded Convolutional Neural Network (MTCNN). The proposed detector replaces selected standard convolutions with depthwise-separable convolutions to reduce computation and parameter count, incorporates Squeeze-and-Excitation (SE) blocks to preserve channel selectivity after compression, and follows an 8-bit integer (INT8)-oriented deployment workflow in which post-training quantization (PTQ) is used for deployment preparation. The optimized detection-oriented cascade is deployed on the AMD Kria KV260, where P-Net, R-Net, and O-Net are implemented as Vitis HLS kernels and integrated through Advanced eXtensible Interface (AXI)-based communication between the processing system (PS) and programmable logic (PL). In this lightweight configuration, O-Net performs final face classification and bounding-box regression only, without the landmark branch of the original full MTCNN. Under the reported evaluation settings, the lightweight model reduces the parameter count from 495,850 to 197,202 (60.2%) and achieves 96.3% AP under the WIDER FACE validation protocol, compared with 97.8% for the original MTCNN. In the current KV260 implementation, the complete deployed system sustains 30 frames per second (FPS), while standalone FPGA kernels at 100 MHz achieve latencies of 0.035 ms, 0.810 ms, and 3.525 ms for P-Net, R-Net, and O-Net, respectively. These results support the feasibility of the proposed framework for real-time edge face detection under resource constraints.