Research On Video Salient Object Detection Based On Deep Neural Network

With the further development in the field of video salient object detection,there are some video salient object detection models that can predict the salient region in the video sequence.However,due to the transformation of complex scenes,the nonsignificant objects motion interference,and the lack of data sets for detecting video salient objects,the existing models need to be improved in the extraction of the overall shape and the boundaries of salient objects.Therefore,aiming at the problem of capturing the overall shape of salient objects and fuzzy boundaries,this paper studies the video salient object detection problem based on deep neural network.The main works are as follows:(1)To accurately capture the overall shape of salient objects in a dynamic scene,this paper proposed a video salient object detection model based on attention feedback network.Which takes the attention feedback network as the backbone of the static saliency module to reduce the loss of visual saliency information caused by scale space problems,and guide the correct integration of multi-scale features from coarse scale to fine scale.Then we extract the multi-scale feature maps extracted from the five decoder blocks of the attention feedback network and transfer them to the pyramid dilated convolution module to retain more spatial visual saliency information.Finally,the Saliency-Shift-Aware conv LSTM is used to integrate the spatiotemporal saliency information and capture the saliency shift.In order to prove the effectiveness of the proposed model,we carried out some ablation experiments on the simple test set of DAVSOD.Experimental results show that the proposed model has achieved good performance in terms of evaluation results and visual effects.(2)In order to obtain the fine boundaries of targets,a novel hybrid loss function is introduced on the basis of the video salient object detection model based on the attention feedback network.This function consists of the predicted Saliency-shift-aware attention map,the prediction results of the salient objects and the salient object boundaries,which is helpful to learn the exact boundaries of the targets.Since no additional prophase or postphase processing is required to optimize the target boundaries,the processing time of each frame is 0.11 seconds.We compared and analyzed the advantages and disadvantages of the proposed model and four state-ofthe-art video salient object detection models on six popular datasets.The experimental results show that the proposed model is superior or equal to the state-of-the-art models in terms of the evaluation results and visual effects.