Dynamic Saliency Detection Based On Deep Learning

Visual saliency is the common feature of Primates.It enables us the ability of process complex images and extract the most valuable regions.The scientist had devoted a lot of energy to research it.They were successfully in combining it with computer science and made it a sub-research direction of computer vision.And helpfully it had done a lot favor for image preprocess in comprehensive vision task.Vision saliency consists of static saliency,co-saliency and dynamic saliency,dynamic saliency is named video saliency as well.Of which the static saliency has been hot researching direction for years and it has achieved outcome good enough.While the dynamic saliency remains to be optimized since it contains plenty work to be done.Since the deep learning networks have draw a lot of attention for its' great performance.It can be introduced to vision saliency detection which help to produce precisely salient object region.Considering the lack of dataset,a novel frame synthesize algorithm has been used for dynamic dataset generating which afford the network a large amount of training data and guarantee the learning module a variety sample space.Category independent object proposal algorithm come with a set of high probability suggest regions.Uniting that kind of regions and the saliency output would finally give the object a better border region.With the improves above,the experiment result also feed back batter.In the end,the algorithm mentioned have be compared with the classic ones.Also,the result shows that our algorithm performs well or even better under certain circumstances.