Msconv: Multiplicative and Subtractive Convolution for Face Recognition

Abstract

In Neural Networks, there are various methods of feature fusion. Different strategies can significantly affect the effectiveness of feature representation, consequently influencing the model's ability to extract representative and discriminative features. In the field of face recognition, traditional feature fusion methods include feature concatenation and feature addition. Recently, various attention mechanism-based fusion strategies have emerged. However, we found that these methods primarily focus on the important features in the image, referred to as salient features in this paper, while neglecting another equally important set of features for image recognition tasks, which we term differential features. In this paper, we attempt to consider both salient features and differential features in the learning process, allowing them to complement each other, thereby achieving more comprehensive and accurate feature learning. We not only propose a novel technique for learning salient features, called Multiplication Operation (MO), but also introduce a novel method named Subtraction Operation (SO) for learning differential features. Based on that, this paper proposes an efficient convolution module called MSConv (Multiplicative and Subtractive Convolution). Specifically, it utilizes multi-scale mixed convolution to extract both local and broader contextual information from facial images, identifying salient features and differential features through Multiplication Operation and Subtraction Operation respectively. Experimental results demonstrate that the MSConv, by learning both salient features and differential features, performs better than models that focus only on salient features.

Keywords: Face recognition, Feature Fusion, Salient Features, Differential Features