Abstract: Facial expression recognition is widely used in many scenes,but the recognition accuracy is reduced due to the influence of illumination,facial appearance and occlusion.In view of the phenomenon,an improved Island loss was proposed to assist facial expression recognition.Although Island loss can improve test performance,it reduces generalization performance.To solve this problem,the improved Island loss used exponential loss to reduce intra-class distance to improve the generalization performance,and used cosine distance loss and Euclidean distance loss to increase inter-class distance to further improve the test performance.The experimental results indicate that LCE loss and ECE loss can both improve the test performance and generalization performance at the same time compared with Island loss.The average generalization performance of ECE loss on CK+,JAFFE,SFEW,RAF and Older datasets increases by 1.05%.