Graph-based Saliency Detection: Integrating Multilevel Features and Analyzing the Distraction Using DNN

Abstract

Salient object detection is a key component in a number of computer vision applications, including scene understanding, object recognition, and image understanding. In salient object detection tasks, deep neural network models have become more popular recently, and have displayed astounding performance. But there is definitely potential for development, especially in terms of capturing multi-level characteristics and dealing with distractions in complicated environments. We suggest a fully convolutional neural network (FCN) with a dilation kernel to resolve the time-consuming shortcoming of conventional CNN, this article seeks to improve salient object detection where one attempt will be made to map the raw input to a dense spatial map and due to this, complex and low-contrast images are also able to give useful information. Distraction analysis finds the parts of an input image that interfere with the saliency detecting process. As a way to determine the saliency value of each node by supplying a coarse saliency map, the graph architecture allows to define the node with various edges. The boundary of the salient item is accurately highlighted using the active contour refinement method. On six well-known public data sets and nine cutting-edge saliency detection techniques, extensive trials are run. In relation to accuracy and robustness, our framework outperforms competitors in a variety of challenging situations, according to the results.