Face detection in complex environments remains challenging due to trade-offs between accuracy and computational efficiency, particularly for edge devices with limited resources. GhostNet-MTCNN is proposed. In this approach, the computationally intensive standard convolution layers in the Multi-task Cascaded Convolutional Neural Network (MTCNN) backbone are replaced by Ghost bottleneck modules from the GhostNet network, which offer lower computational requirements. This modification reconstructs the network’s feature extraction capabilities, resulting in a new model. Experimental results demonstrate that the proposed method effectively balances model parameters and accuracy. On the Easy, Medium, and Hard validation sets, the method improves accuracy by 5.6%, 6.6%, and 7.8%, respectively, with only a 0.62M increase in parameters compared to MTCNN. In comparison to MobileNetV3-MTCNN, the model reduces the parameter count by 1.27M while improving accuracy by 1.6%, 0.8%, and 0.5%, respectively. Hardware deployment on FPGA further validates its practical efficacy, achieving a $400\times $ acceleration in convolution operations through optimized parallelization and memory caching. This study enhances the MODEL’s ability to detect small-size, multi-angle faces in low light, partially occluded, and noisy environments in complex scenarios, while effectively balancing parameter count and detection accuracy, making it a superior choice for deployment on edge devices.