Research On Video Saliency Detection Algorithm Based On Attention Mechanism

Saliency object detection aims to identify the areas that are most important and interesting to the human eyes.As a popular direction in the field of computer vision,it is currently mainly used in target segmentation,motion recognition,target tracking and other tasks.The attention mecha nism of human vision means that people can allocate more attention to salient regions in complex scenes,which provides a feasible solution for salient object detection.Imitating the attention mechanism of human vision and aiming at the problem of no difference between the processing characteristics of neural networks in the current video saliency detection research,this paper refers to the attention mechanism to improve the attention degree to salient areas and completes the design of video saliency detection algorithm based on spatial domain attention mechanism.At the same time,according to the inconsistency of time and space information in video feature extraction,this paper designs a conv LSTM network structure embedded in the attention mechanism,so that the network can extract spatiotemporal consistency information while maintaining the spatial structure information of features.And finally we design a video salient object detection model based on attention mechanism.Aiming at the problems of video saliency detection,the main research contents of this article are as follows:(1)Aiming at the difference between features and different contributions to the prediction of saliency targets,the attention mechanism is introduced to distinguish features,and a video saliency detection algorithm based on spatial domain attention mechanism is designed.The core of the algorithm is to apply the attention mechanism based on channel-wise and spatial-wise attention to video saliency detection,so that the network can pay more attention to the saliency area.The algorithm considers that the differences in the characteristics of different layers of the network: the higher layers of the network extract more semantic features,and the shallower layers of the network pay more attention to detailed information.The CA unit based on channel-wise attention is added to the higher layer of the network to improve the network's ability to locate saliency areas.Add SA units based on spatial-wise attention to the shallower layers of the network to improve the network's ability to suppress background noise.At the same time,we design high-level features to guide the learning of shallow features to make the network better capture salient objects.Finally,the ablation experiment is conducted to verify the effectiveness of the algorithm innovation point through F-measure value,MAE value and other indicators.(2)For the problem of how to model the information in the time domain for video saliency detection,this paper designed a attention-based spatiotemporal consistency video saliency detection algorithm in order to perform more efficient spatiotemporal feature prediction.The algorithm has designed a conv LSTM module embedded in the attention mechanism at the higher layer of the network,embedding the channel-wise and special-wise attention mechanism into the conv LSTM structure to obtain more accurate spatiotemporal consistency information.A two-layer conv LSTM network unit is used to model the correlation between the front and back frames of the video frame sequence.Finally,through the ablation experiment,the effectiveness of the algorithm innovation is verified.At the same time,the effect of the number of frames on the network performance is verified by changing the number of video frame sequences.(3)This paper verifies the advancedness of the algorithm we proposed through sufficient experiments.We compare 11 mainstream algorithms based on traditional methods and 6 deep learning methods in the video saliency detection field,and conducte experimental verification on 6 recognized video datasets such as DAVIS,FBMS,and MCL.From a qualitative perspective,we can see that compared with other mainstream algorithms,our algorithm has the advantages of accurate positioning and clear detection of the target edge.From the quantitative indicators,we can see that our algorithm have achieved good results in a recognized evaluation index such as F-measure value,MAE value.We adopted the end-to-end learning method of the neural network without any pre-processing or post-processing,and finally reached a processing speed of 20 fps.