Convolutional Neural Networks Hyperparameter Tunning for Classifying Firearms on Images

Figures & data

Figure 1. Estimated breakdown of intentional homicide worldwide, by a mechanism of perpetration in 2017 UNODC (2019)

Table 1. Works that propose CNN architectures for detecting fire guns in images.

Reference	Gun Type	No. of Images	CNN Arch.	Transfer Learning	Hyper parameters	Accuracy
Kaya, Tuncer, and Baran (2021)	assault rifles	5214 weapon images	VGGNet,	No	ReLU activation function	VGGNet 98.40%
	bazookas, grenades	from the internet	ResNet-50		32 mini batch size	VGG-16
	hunting rifles, knives,		ResNet-101		0.25 dropout	89.75%
	pistols and revolvers				Adamax optimization alg.	ResNet-101 83.33%
					30 epochs
Ağdaş, Türkoğlu, and Gülseçen (2021)	gun and knife images	16000 images containing	AlexNet	Yes	Mini-batch size 8–16	Alexnet 97.74
		9500 knives, 3500 guns,	VGG16,	fine tuning	Maximum epoch number 5–20	VGG16 99.38
		and 3000 ordinary pictures	and VGG19		Weight decay factor 0.01	VGG19 99.27
		Images from the Internet			Initial learning rate 0.0001
					Optimization method SGDM
					(Stochastic Gradient Descent
					with Momentum)
Dwivedi, Kumar Singh, and Singh Kushwaha (2019)	gun and knife	Training set contains	VGG-16	Yes,	Model A dropout 0.5	Model A
		1520, 1800 and 1176		fine tuning	Model B dropout 0.7	obtains
		images of knives, guns			Batch size 8,	98.41%
		and no-weapons, respectively.			and 20 epochs for	accuracy
		Test set contains 344, 368			both models	(acc.)
		and 296 images of same
		classes as the train set
Veranyurt and Okan Sakar (2020)	images with	Gun Data Set has	Basic CNN	Yes,	Optimization method	VGG16 with
	and without	3000 images	from scratch	fine tuning	SGDM.	transfer learning
	a hand-gun	Film Gun Data Set	and with		12 epochs	88% of accuracy.
		has 333 images	transfer learning			VGG16 with
		Coco 2017 Validation Data Set	VGG-16 from			transfer learning
		has 10,000 images	scratch and with			95% of accuracy
			transfer meaning			with 12 epochs
Olmos, Tabik, and Herrera (2018a)	pistols, machine	3000 images of guns	VGG-16	Yes,	Optimization	The best
	guns, rifle,	taken from a variety		fine tuning	method SGD	obtained accuracy
(Stochastic Gradient Descent)	grenade, rocket	of contexts. Authors created			(Stochastic Gradient Descent)	was 94.44%
	launcher, tank	four databases			with back-propagation
		ImageNet for Transfer Learning
Elmir, Ahmed Laouar, and Hamdaoui (2019)	handgun	Gun Data Set consist	MobileNet	No	Not specified	MobileNet 90% acc.
		of 3000 images	CNN basic			CNN basic 55% acc.
			Fast R-CNN			Fast R-CNN 80% acc.

Figure 2. Basic architecture of a CNN.

Figure 3. Examples of the contents of our data set for both categories.

Table 2. Results using default hyperparameter values for Keras optimizers for both AlexNet and Inception V3.

CNN	Optimizer	Batch size	Learning rate	Epochs	True pos	True neg	Accuracy
AlexNet	Adadelta	32	0.001	200	240/450	418/450	73.11%
”	Adagrad	32	0.001	200	255/450	409/450	73.78%
”	Adamax	32	0.001	200	285/450	353/450	70.89%
”	Nadam	32	0.001	200	238/450	402/450	71.11%
”	SGD	32	0.01	200	278/450	402/450	75.56%
Inception	Adadelta	32	0.001	200	377/450	443/450	91.11%
V3	Adagrad	32	0.001	200	375/450	444/450	91.00%
”	Adamax	32	0.001	200	304/450	444/450	83.11%
”	Nadam	32	0.001	200	299/450	415/450	79.33%
”	SGD	32	0.01	200	394/450	442/450	92.89%

Table 3. Results using tuned hyperparameter values for Keras optimizers for both AlexNet and Inception V3, with its accuracy gain respect to the default values.

CNN	Optimizer	Batch size	Learning rate	Epochs Epochs	True pos	True neg	Accuracy	Change
AlexNet	Adadelta	32	0.01	200	235/450	421/450	72.89%	−0.22%
”	Adagrad	128	0.001	200	239/450	413/450	72.44%	−1.34%
”	Adamax	64	0.0001	200	276/450	409/450	76.11%	+5.22%
”	Nadam	64	0.0001	200	282/450	411/450	77.00%	+5.89%
”	SGD	128	0.01	200	272/450	409/450	75.67%	+0.11
Inception	Adadelta	64	0.01	200	386/450	443/450	92.11%	+1.00%
V3	Adagrad	64	0.01	200	397/450	444/450	93.44%	+2.44%
”	Adamax	128	0.0001	200	394/450	444/450	93.11%	+10.00%
”	Nadam	256	0.0001	200	389/450	444/450	92.56%	+13.23%
”	SGD	128	0.01	200	383/450	441/450	91.56%	−1.33%

Table 4. Confusion matrix for Inception V3 with optimum parameters.

Actual/Predicted	Positive	Negative
Positive	406	44
Negative	9	441

Figure 4. Validation loss and accuracy across epochs during the training process of AlexNet network with tuned hyperparameters.

Table 5. Confusion matrix for AlexNet with optimum hyperparameters.

Actual/Predicted	Positive	Negative
Positive	282	168
Negative	39	411

Figure 5. Validation loss and accuracy across epochs during the training process of Inception V3 network with tuned hyperparameters.

Figure 6. ROC curve for AlexNet network with tuned hyperparameters. Its area under the curve value is 0.8709.

Figure 7. ROC curve for Inception V3 network with tuned hyperparameters. Its area under the curve value is 0.9812.

Figure 8. Examples of misclassification.

UNODC. 2019. “Global study on homicide 2019.” July. Accessed 2019. https://www.unodc.org/unodc/en/data-and-analysis/global-study-on-homicide.html.

 Google Scholar

Kaya, V., S. Tuncer, and A. Baran. 2021. Detection and classification of different weapon types using deep learning. Applied Sciences 11 (16):7535. doi:10.3390/app11167535.

 Google Scholar

Ağdaş, M. T., M. Türkoğlu, and S. Gülseçen. 2021. Deep neural networks based on transfer learning approaches to classification of gun and knife images. Sakarya University Journal of Computer and Information Sciences 4 (1):131–2784. doi:10.35377/saucis.04.01.891308.

 Google Scholar

Dwivedi, N., D. Kumar Singh, and D. Singh Kushwaha. 2019 April. “Weapon Classification using Deep Convolutional Neural Network.” In 2019 IEEE Conference on Information and Communication Technology, pp.1–5. Manhattan, New York, U.S.: IEEE. doi: 10.1109/CICT48419.2019.9066227.

 Google Scholar

Veranyurt, O., and C. Okan Sakar. 2020. “Hand-Gun detection in images with transfer learning-based convolutional neural networks.” In 28th Signal Processing and Communications Applications Conference (SIU), Gaziantep, Turkey, January, pp. 1–4. Manhattan, New York, U.S.: IEEE. doi: 10.1109/SIU49456.2020.9302394.

 Google Scholar

Olmos, R., S. Tabik, and F. Herrera. 2018a. Automatic handgun detection alarm in videos using deep learning. Neurocomputing 275:66–72. 2017.05.012. doi:10.1016/j.neucom.

 Web of Science ®Google Scholar

Elmir, Y., S. Ahmed Laouar, and L. Hamdaoui. 2019. “Deep learning for automatic detection of handguns in video sequences.” In JERI, April. http://ceur-ws.org/Vol-2351/paper_69.pdf.

 Google Scholar