Improved Loss Function for Image Classification

Table 2

Comparison of recall rates with three loss functions.

Dataset	Model	Loss function	Recall@1	Recall@2	Recall@4	Recall@8
CIFAR-100	GoogLeNet	Margin loss	0.607	0.712	0.763	0.815
		Triplet loss	0.610	0.708	0.772	0.810
		New loss	0.632	0.720	0.779	0.817
	GoogLeNet	Margin loss	0.712	0.810	0.873	0.916
		Triplet loss	0.695	0.801	0.869	0.912
		New loss	0.722	0.8291	0.8892	0.9294
CIFAR-10	ResNet50	Margin loss	0.7807	0.8496	0.8922	0.9312
		Triplet loss	0.8067	0.8661	0.9045	0.9319
		New loss	0.7743	0.8638	0.918	0.9504
Fashion-MNIST	ResNet50	Margin loss	0.8864	0.9273	0.9545	0.9704
		Triplet loss	0.8909	0.9304	0.959	0.9742
		New loss	0.8749	0.9259	0.961	0.9801