Salient Object Detection Based On Deep Learning

Salient object detection is an important first step for the computer to understand the surrounding environment.Its purpose is to allow the computer to imitate the human attention mechanism to detect attractive areas in the image.These attractive areas contain most of the visual information of the image.By filtering out the foreground areas of the image that contain the main visual information,the subsequent steps of image understanding can not only obtain cleaner and more accurate content information in the image but also reduce the resource of calculation and storage when processing the background area of the image,thereby improving the overall performance of image understanding 's subsequent steps.Therefore,salient object detection is widely used in downstream computer vision tasks,such as image and video compression,image segmentation,image recognition,image synthesis,image search,etc.In the research field of computer vision,salient object detection can be defined as a binary segmentation problem.At present,although many salient object detection models based on traditional machine learning methods and deep learning have been proposed,the results of these models still have two deficiencies: The one is that the salient objects can be kinds of objects with any shapes,so the boundaries of the salient objects have odd shapes without obvious rules to follow,and it is difficult for the convolutional neural network to segment the boundary of an object with arbitrary shape.Therefore,most of the current salient object detection models can roughly locate the position of salient objects,but the boundaries are relatively blurred.The other is that most salient object detection models rely on a large number of labeled images for training,but obtaining the labels of training images requires a lot of manpower,material resources,financial resources and time.This paper proposes two models to solve the two problems mentioned above.They are “Saliency Detection based on holistic and deep feature pyramids” and ”Semi-supervised salient object detection model based on two concatenated generative adversarial networks”.Through a large number of experiments,this paper shows that the “Saliency Detection based on holistic and deep feature pyramids” can effectively alleviate the problem of blurring the boundary of the existing saliency detection model from the network structure design.At the same time,the “Semi-supervised salient object detection model based on two concatenated generative adversarial networks” can significantly reduce the labeled training data required by the saliency detection model through semisupervised training.This provides the possibility that the saliency detection model can be applied to actual application scenarios which have limited labeled data.