In recent years, facial expression recognition (FER) has garnered significant attention within the realm of computer vision research. This paper presents an innovative network called the Dual-Direction Attention Mixed Feature Network (DDAMFN) specifically designed for FER, boasting both robustness and lightweight characteristics. The network architecture comprises two primary components: the Mixed Feature Network (MFN) serving as the backbone, and the Dual-Direction Attention Network (DDAN) functioning as the head. To enhance the network’s capability in the MFN, resilient features are extracted by utilizing mixed-size kernels. Additionally, a new Dual-Direction Attention (DDA) head that generates attention maps in two orientations is proposed, enabling the model to capture long-range dependencies effectively. To further improve the accuracy, a novel attention loss mechanism for the DDAN is introduced with different heads focusing on distinct areas of the input. Experimental evaluations on several widely used public datasets, including AffectNet, RAF-DB, and FERPlus, demonstrate the superiority of the DDAMFN compared to other existing models, which establishes that the DDAMFN as the state-of-the-art model in the field of FER.

Optical neural networks (ONNs) are getting more and more attention due to their advantages such as high-speed and low power consumption. However, in a non-ideal environment, the noise and low-bits control may heavily lead to a decrease in the accuracy of ONNs. Since there is AD/DA conversion in a simulated neural network, it needs to be quantified in the model. In this paper, we propose a quantitative method to adapt ONN to a non-ideal environment with fixed-point transmission, based on the new chip structure we designed previously. An MNIST hand-written data set was used to test and simulate the model we established. The experimental results showed that the quantization-noise model we established has a good performance, for which the accuracy was up to about 96%. Compared with the electrical method, the proposed quantization method can effectively solve the non-ideal ONN problem.