A new frog leaping algorithm-oriented fully convolutional neural network for dance motion object saliency detection

Image saliency detection is an important research topic in the field of computer vision. With traditional models, texture details are not obvious and edge contour complete. The accuracy recall rate object low, which mostly based on manual features prior information. rise deep convolutional neural networks, has been rapidly developed. However, existing methods still have some common shortcomings, it difficult to uniformly highlight clear boundary internal region whole complex images, mainly because lack sufficient rich features. In this paper, a new frog leaping algorithm-oriented fully network proposed for dance motion detection. VGG (Visual Geometry Group) model improved. final full connection layer removed, jump used prediction, can effectively combine multi-scale information from different convolution layers network. Meanwhile, improved algorithm optimize selection initial weights during initialization. process iteration, forward propagation loss calculated, anomaly weight corrected by using algorithm. When satisfies terminal conditions, optimized one make further optimization. addition, high-level semantic low-level detail data-driven framework. order preserve unity inner effectively, connected conditional random (CRF) adjust obtained feature map. precision (PR) curve, F-measure, maximum weighted F-measure mean absolute error (MAE) tested six widely public data sets. Compared with other most advanced representative methods, results show that method achieves better performance superior methods. presented reveals strong robustness image various scenes, more uniform accurate.