Gun Dataset: http://www.mediafire.com/file/pvfircmboaelkxc/Gun_data_labeled.zip/file
Knife Dataset: http://kt.agh.edu.pl/matiolanski/KnivesImagesDatabase/
we trained the model on VGG16 architecture. The proposed model gave 97% accuracy on the validation dataset. The models proposed in previous papers gave an accuracy of up to 87%. The proposed system can predict the firearm with an accuracy of 97%. The proposed system can be used in surveillance cameras for detecting the gun. The surveillance cameras are used in a variety of locations, such as malls, cinema theaters, buildings, etc. In the VGG16 model, there are 135 million parameters to be trained for detecting the gun in an image. We used ‘ReLU’ as an activation function in the convolutional layers and ‘softmax’ as an activation function in the output layer. We used Adam (lr=0.0001) as an optimization function. We used ‘categorical cross-entropy’ as a loss function. We trained the model up to 10 epochs.
ResNet arcitechure and the proposed model gave 95.6% accuracy on a validation dataset. The ResNet model can be used in surveillance cameras for detecting firearms. In the ResNet model, there are 23.5 million parameters to be trained and 53 thousand non-trainable parameters. We used ‘softmax’ as an activation function in the output layer. We used Adam (lr=0.001) as an optimization function. We used ‘categorical cross-entropy’ as a loss function. We trained the model up to 30 epochs. We added one GlobalAveragePooling2D layer and a dropout (0.7) before the output layer to avoid overfitting
ResNet Archietcure: MobileNet model was proposed with the aid of google. The MobileNet model was proposed such that it can be used on an embedded system with less processing powers and memory. The proposed model gave 99% accuracy on a validation dataset. The MobileNet model can be used in surveillance cameras, embedded devices, etc.. In the MobileNet model, there are 28.8 million parameters to be trained and 22 thousand non-trainable parameters. We used ‘ReLU’ as an activation function in the convolutional layers and ‘softmax’ as an activation function in the output layer. We used Adam (lr=0.001) as an optimization function. We used ‘categorical cross-entropy’ as a loss function. We trained the model up to 40 epochs. We added one more dense layer before the output layer.