In the past few years, due to its highly beneficial applications, the exploration of violence detection has become an increasingly pertinent subject in the realm of computer vision, leading to the development of numerous proposed solutions by researchers. This work presents a Transfer learning approach for violent scene detection, with redundant frames removal from video frames for classification into violent and non-violent categories. Two benchmark datasets are used for validating the proposed technique for abnormal event’s detection; inception v3,

a state-of-the-art Convolutional Neural Network (CNN) composed of 316 layers is used. The model is pre-trained with a benchmark ImageNet dataset’ which includes 1000 classes. In the first phase, redundant video frames from the dataset are removed using an unsupervised learning approach. To detect whether the two frames are similar or not; features from both frames are extracted, and Euclidean distance is computed. If the distance between the frames is small, then the image is considered redundant and is dropped from the dataset. In the second phase, transfer learning is performed to detect violent frames. In transfer learning, the CNN model is updated by removing the last three layers and adding a fresh layer to the network. Hold out cross-validation technique is applied to partition the dataset into train and test sets, data is randomly divided into 70- 30 percent where 70% data is selected for training and 30% is selected for testing. The redundant frame removal method helps in efficient training and validation by reducing the size of the dataset. The proposed model performed much better than state-of-the-art approaches and achieved an accuracy of 78.61% for the real-life violence dataset and 73.86% for the Hockey Fight Dataset. The accuracy can be further increased by choosing the best optimal threshold value for key frame selection and tuning the epoch and other hyper parameters.

Index Terms- Computer Vision, Automatic Violence Detection

System, Convolution-al Neural Networks (CNN), Transfer

Learning, Action Recognition.