LAUSR

Recently, significant progress has been made in the field of face detection. However, despite the demand for its high accuracy and recall rate, the efficiency of face detection algorithm is another key factor in evaluating its performance, which puts forward serious challenges to current models. To boost up the efficiency of face recognition, in this paper, we propose a lightweight rapid framework, called LRNet, which has fewer convolutional layers and a higher efficiency. In particular, our framework is consist of two major modules. One is the Feature Map Fast Shrink Module (FMFSM), which leads to a fast reduction in the size of feature maps and time consumed for detection by using Two Information Flow Block (TIFB). Another module, namely Variable Scale Face Detection Module (VSFDM), is consist of the Retinal Receptive Field Block (RRFB) and designed to prevent a single or composite feature map from undertaking too much tasks. In addition, we propose a new anchor strategy that considers not only the density of anchors with different scales but also the position and central symmetry of the features. Our proposed LRNet achieves high accuracy and efficiency on the challenging FDDB dataset for face detection. When the number of false positives is 2000, its True Positive Rate (TPR) under discrete and continuous scores can achieve 0.951 and 0.725 respectively. When running on GTX 1080Ti, given images with a resolution of $1024\times 1024$ , the average time consumed for detection is merely 8.88 ms.