Holistic Attention on Pooling Based Cascaded Partial Decoder for Real- Time Salient Object Detection

Abstract

The identification of saliency is one of the fundamental computer vision problems. Convolutional Neural Network (CNN) has recently become the most commonly employed and potent detection strategy in which feature maps from various layers are combined without distinction. The current salient object detection frameworks use the multi-level aggregation of pre-trained neural networks. We resolve saliency identification via a cascaded partial decoder convolutional neural network with a holistic attention framework while focusing on extending the pooling function. Our framework is a partial decoder that discards relatively large resolution features for the acceleration of shallow layers. On the other hand, we noticed that deeper layers feature integration to achieve a reasonably accurate saliency map. We employed a pooling-based module based on a U-shaped architecture with a holistic attention model to gradually optimize the highly-developed semantic features of the backbone network. Experimental findings on six benchmark data sets show that we are more reliably locating the salient objects with sensitive information, thus improving their accuracy considerably compared to 22 prior art. Our method is as fast as it does to process an image of size 352×352 at a pace more significant than the different state-of-the-art architectures.