Research on image recognition method based on LMAL and VGG-16
Ying Shan Cao, Runlong Gu, Runlong Gu, Chenghua Huang · 2022
In deep convolutional neural networks, the traditional Softmax loss function lacks the ability to distinguish similar classes. In order to solve this problem, the idea of increasing the inter-class spacing and reducing the intra-class spacing is widely recognized. The large margin angular loss (LMAL) loss function is introduced to reduce the intra-class spacing by L2-standardization of the features and weight vectors of the Softmax loss function. At the same time, LMAL also has a good ability to distinguish deeper features. Combined the LMAL loss function and the VGG-16 model, the results on three independent datasets show that the image recognition accuracy of the improved model has been significantly improved.