Multiscale Strategy For Visual Saliency Detection

The essence of saliency detection is to simulate the focus of human eyes using some learnable models or prior knowledge,which is mainly applied to the early processing of other computer vision tasks for providing them with early acceleration processing such as image compression,video compression and camera autofocus.Currently,the existing Multiscale Deep Features detection(MDF)algorithm uses the depth network for feature vector extraction for many times,and scales the second-level super pixel block and global image composed of super pixel block,adjacent super pixel block to 227×227 size,and then uses the depth network to extract high-dimensional feature vectors,resulting in the problem of redundant computing.In this paper,an acceleration strategy for efficient computation is proposed.Based on the limited feature information contained in general small-scale image blocks,this paper scales the area and global image composed of super-pixel blocks,adjacent super-pixel blocks to different sizes,and then merges them into a 227×227 image for high-dimensional feature vector extraction.At first,the raw image is segmented into many super pixels by using graph-based segmentation technology(EGBIS),and in order to resist the influence of hyperparameters of super pixels segmentation,15 groups of different hyperparameters are used to segment the image to have multi-level and multi-scale information.Then iterate through all the super pixel block,to extract the current regional block and composed of adjacent secondary image pixel block and a global image,the three dimension,and respectively scaling to a different certain size,for merging into a 227 × 227 the size of the image.And then feeding the merged map to AlexNet network for high-dimensional feature extraction to train a senior decision to classify the current superpixel whether saliency or not.At last,a set of fusion parameters is trained from the 15 primary significance images,and the fusion results into an optimal one.The work of this paper mainly includes the following four aspects:(1)The existing algorithm exists redundant computation,because multiple scaling of different scales to the same size and then in turn network computing.A kind of to use15 different set super parameter image segmentation technology(EGBIS)is proposed to segment MSRA-B every image data sets to pixels of segmentation,traverse each segmentation results on the graph of each pixel block,and the adjacent pixel blocks of area,and the scale of the global image,will be zoomed to the different fixed size,and to merge into one picture,feeding it to the deep neural network to determine whether is salient.(2)Because the data set samples required by the algorithm proposed in this paper are different from the distribution of images trained by the basic network AlexNet,two new all-connected layers are added based on the basic network AlexNet,and the number of neurons in each layer is 300.At the same time,gaussian distribution is adopted for the initialization of the parameters of the new full connection layer.The parameters of the convolutional layer set by the basic network AlexNet are fixed and not updated,and the parameters on the full connection layer are open for secondary training and update.(3)After the deep neural network produces 15 primary significance graphs,this paper proposes to train a group of parameters to fuse into the optimal significance graph.The optimal result graph is learned through the value of the minimum binary cross entropy loss function.(4)Our improved deep model and the existing MDF algorithm were used to calculate PR curve,F-measure and mean absolute error MAE on four publicly available MSRA-B test sets,PASCAL-S,HKU-IS and ECSSD data sets.The experimental results of the multi-scale detection method of visual significance show that the algorithm proposed in this paper is better than MDF in the test set of msra-b,and almost as good as MDF algorithm in the other three data sets.In addition,the calculation time required for the test of the same size image 300×400 is 3.32 s,which is obviously lower than the calculation result of MDF algorithm 8.0s.