Abstract:  In recent years, face detection based on convolutional neural networks (CNN) has dominated this field, and
the detection results on the public benchmark set have also been significantly improved. However, the computational
cost and model complexity are on the rise. It remains a challenge to apply face detection model to embedded devices
with limited computing power and memory capacity. Aiming at the application of face detection of 320×240
resolution input images in embedded systems, a low-resolution face detection algorithm based on lightweight network
was proposed. The backbone network employed the attention module, combined Distance-IoU (DIoU) and Non-Maximum Suppression (NMS), and adopted the Mish activation function. Meanwhile, an appropriate a priori box
was set for the face feature ratio. In doing so, the balance could be achieved between precision and speed, and it could
be deployed to the embedded platform. Specifically, deep separable convolution was used to replace ordinary
convolution, and an attention convolutional block attention module (CBAM) was added after the convolution block to
keep the network’s focus on the target object to be recognized. Instead of the ReLU activation function, the Mish
activation function was used to improve the model inference speed. By combining DIoU and NMS, the algorithm’s
detection accuracy for small faces was enhanced. The results of experiments on the WIDER FACE dataset prove that
the proposed method not only can detect human faces with high accuracy in real time, but also has higher accuracy
than traditional algorithms in small resolution input. After expanding the dataset, the proposed model also improves
the detection accuracy under complex illuminations.

Key words: face detection, lightweight network, attention module, activation function, non-maximum suppression