An Enhanced Lightweight Convolutional Cascade Network Based on MobileNetV3 and MTCNN for Robust Face Detection in Complex Environments

Abstract

In order to improve the detection accuracy of small-size and multi-angle faces in complex environments while maintaining efficiency for deployment on edge devices, this paper proposes an enhanced neural network model—MobileNetV3-MTCNN. Based on the classical MTCNN architecture, the proposed model reconstructs the feature extraction network by integrating the MobileNetV3 Bottleneck module, optimizing the balance between computational efficiency and detection precision. The improved architecture enables more effective multi-scale feature extraction and reduces redundancy in network operations. Experimental results on the WIDER_FACE dataset demonstrate that MobileNetV3-MTCNN achieves substantial gains in average precision (AP) and recall rate across easy, medium, and hard subsets compared to the original MTCNN, with increases of 4.8%, 6.2%, and 7.5% in AP respectively. Although the model parameters slightly increase, the proposed method demonstrates superior robustness and adaptability in detecting small and occluded faces under diverse lighting and pose conditions. The study confirms that the MobileNetV3-MTCNN model provides an effective and lightweight solution for real-world face detection applications, especially in densely populated and dynamic environments.