A transfer learning-based deep focal multiclass network for psychological emotion recognition in community-correction populations

Facial expression-based emotion recognition technology holds significant application value in fields such as intelligent security and psychological intervention. In particular, for individuals under community correction, automated emotion analysis can assist in psychological assessment and behavioral risk monitoring, thereby enhancing the scientific rigor and real-time effectiveness of interventions. Recent studies have proposed various deep learning-based classification models to improve emotion recognition performance in complex scenarios. However, the performance of psychological emotion recognition models based on facial images is still limited by factors such as the scale of training data and class imbalance. To address these challenges, this study focuses on the task of psychological emotion recognition for community-correction populations and propose a novel Transfer Learning-based deep Focal multiclass Network (TLFNet). Specifically, the TLFNet model incorporates a new multiclass Focal Loss function to optimize its parameters, which enhances the model’s sensitivity to minority-class samples and mitigates the bias introduced by class imbalance. Moreover, under the transfer learning framework, TLFNet adopts ImageNet pre-trained weights to incorporate large-scale visual prior knowledge. Extensive experiments conducted on a real-world emotion recognition dataset demonstrate the effectiveness of each component of the TLFNet model and further validate its superior overall performance in the target task.