Research On The Method Of Saliency Object Detection Based On Deep Learning

With the rapid popularity of Internet technology and digital products, image has become an important information carrier in people's life by changing people's way of life in each aspect. And this also means that how to efficient processing vast amounts of information, has become an important research topic. When the computer processing the images, at the same time to deal with all the details of image information is very inefficient. If the limited computing resources allocated to the target area of the image, is a very effective way to improve the computational efficiency directly. Salient region detection in recent years, therefore, become a very important research direction in the field of computer vision. Through the images of features such as color, texture, location analysis, so as to accurately locate and segment the target, is significant to detect the main goal. With machine learning, deep learning algorithms such as matures, saliency detection areas are also usher in a new development.This paper is an attempt to use deep learning into salient region detection, proposed a new saliency detection algorithms based on deep learning.After setting up a platform for deep learning, We modify and configure SPP code which was used in mage semantic segmentation,then trained a new deep learning model for saliency detection.Using this model,we can get the saliency information of echa picture.These information were showed by boundingboxes and scores.Next stop.we use these information to generate two original saliency maps based two different strategies.Then we extract deep learning feature from the 5th layer of the deep learning model to generate the 3rd original saliency by calculate region contrast with image segmentation results.At last,we use cellular automata to optimize the three original saliency maps and fusion for a final one.In order to fairly evaluate our proposed algorithm,we use MSRA5000 and PASCAL1500 to take the test.We use 3000 of the picture of MSRA5000 to train the model of deep learning,the last 2000 picture were used to take the test.After that, we take a transverse comparison of seven mainstream methods,by using PR curve, ROC curve, and MAE standard. The results show that our algorithm is superior to other algorithms in performance.