An Enhanced Violence Detection Using Convolution Neural Networks Approach

Abstract

Violence detecting through cameras is to find specific object identification and human action is prominent in office sectors and public places to enhance safety. CCTV images contain different objects, different backgrounds, real-time incident in office sectors and public places, in the existing methods residual network and K-nearest neighbors’ method were used to classify and detect violence. The disadvantages of these methods are finding objects and human behavior.  The accuracy level is also the most complex to achieve. Therefore, in the proposed method, violence detection and object identification using the Inception-v3 architecture and networking techniques are used to improve the accuracy.  The Inception-v3 architecture is a convolutional neural network (CNN) that uses a variety of techniques to improve image classification performance. The networking techniques used in research consist of five stages: video input, image augmentation, image classification, email notification, and sound alarm. The video clips from various open-access databases are utilized and converted with different data sets of testing images. Image augmentation improves the blur and saturation of each testing image and makes the testing images better. Yolo v3 algorithm separates the different objects, like the hand, table, and specific parts, for detecting objects. The Inception-v3 architecture classifies each human action for the identification of human behavior. The training images classify each frame of the image dataset, which employs deep learning methods with computer vision technologies. The experiments on these database results with 99.15% accuracy, along with email notification alerts and a sound alarm.