Research On Image Classification And Abnormal Event Detection Algorithm Based On Multi-instance Deep Feature Learning

Economic and social development is changing people's lives.Science and technology has gradually penetrated into all aspects of society.With the developing of social media,People prefer to use pictures to express their opinions.The small and flexible monitoring equipment has also improved the level of social security.In recent years,multi-instance learning has been widely used in the multimedia retrieval tasks,for its unique bag-instance structure can well model the multimedia content.This paper explores multimedia retrieval with combination of multi-instance learning and deep representation learning.The main contents are as follows:(1)To solve the problems that the representation of Image bag is mostly limited to low-level manual features and there are redundant and noisy instances in the bag,a multiinstance algorithm based on deep feature selection is proposed and applied to image classification.The method constructs a multi-instance structure by treating the entire picture as a bag in multi-instance learning,and several semantic regions after image segmentation as instances in the bag.First,the high-level instance representation is learned from a transfer learning model and then the deep bag representation is extracted through mapping to all instances.To eliminate the influence of the noisy instances,an efficient feature selection algorithm is introduced to get more discriminative bag representations.Finally,a learned SVM is used to predict the image bag label.The effectiveness of the algorithm is verified through experiments on the image data set using for the multi-instance algorithm.The proposed method obtains average 10.61% increase in precision.(2)In order to solve the problem that abnormal events in safety surveillance videos are relatively small,difficult to label,and features are easily weakened,this paper proposes a deep network based on weakly supervised learning to apply to video anomaly event detection.The method firstly divides the video into several small time segments.Each segment is treated as an instance and the whole video is treated as a bag.The video containing abnormal segments is considered as a positive bag,otherwise it is considered as a negative bag.Then use the bag as a unit to perform feature learning on the video through the pseudo three-dimensional convolution residual network to obtain the temporal and spatial features of the video,and then use the multi-instances ranking to obtain the abnormal score of the instance,and redistribute the video feature weights through the attention model.Finally,the abnormal score of the video segment is weighted to obtain the abnormal score of the video.The Experiment results on the large anomaly detection data sets UCF-Crime and Shanghai Tech show that the proposed method is 9.39% and 4.82% better than benchmark method.