References	Classification or detectiona	Deep CNN architecture	Training strategyb	Best accuracy(%)c	Evaluation qualityd
Atabay, 2017	C	VGG16, 19, custom architecture	FS–TL	97.53	**
Barbedo, 2018b	C	GoogleNet	TL	87	*
Brahimi et al., 2017	C	AlexNet, GoogleNet	FS–TL	99.18	*
Brahimi et al., 2018	C	AlexNet, DenseNet169, Inception v3, ResNet34, SqueezeNet1-1.1, VGG13	FS -TL	99.76	*
Cruz et al., 2017	C	LeNet	TL	98.60	*
DeChant et al., 2017	D	Custom three stages architecture	FS	96.70	**
Ferentinos, 2018	C	AlexNet, AlexNetOWTBn, GoogleNet, Overfeat, VGG	Unspecified	99.53	*
Fuentes et al., 2017	D	AlexNet, ZFNet, GoogleNet, VGG16, ResNet50, 101, ResNetXt-101	TL	85.98	**
Fuentes et al., 2018	D	Custom architecture with Refinement Filter Bank	TL	96.25	**
Liu B. et al., 2017	C	AlexNet, GoogleNet, ResNet 20, VGG 16 and custom architecture	FS -TL	97.62	*
Mohanty et al., 2016	C	AlexNet, GoogleNet	FS–TL	31	***
Oppenheim and Shani, 2017	C	VGG	Unspecified	96	*
Picon et al., 2018	C	Custom ResNet50, Resnet50	TL	97	***
Ramcharan et al., 2017	C	Inception V3	TL	93	**
Sladojevic et al., 2016	C	CaffeNet	TL	96.3	*
Too et al., 2018	C	Inception V4, VGG 16, ResNet 50, 101 and 152, DenseNet 121	TL	99.75	**
Wang et al., 2017	C	VGG16, 19, Inception-V3, ResNet50	TL	90.40	*
Zhang S. et al., 2018	C	Custom Three Channels CNN, DNN, LeNet-5, GoogleNet	FS	A/ 87.15 B/ 91.16	A/ * B/ *
Zhang K. et al., 2018	C	AlexNet, GoogleNet, ResNet	TL	97.28	*

^aClassification (C)—Detection (D).

^bFrom Scratch (FS)—Transfer Learning (TL).

^cIf available, the accuracy of the explicitly different test set is privileged.

^dSSAbsence of three explicit subsets; SSThree explicit subsets; SSSTest set explicitly different from the training set.