Saliency-Subitizing Computation And Salient Object Detection In Images Based On Deep Learning

Salient object detection in images has drawn considerable attention in computer vision,image processing and other fields,which has become one of the important research branches.Unlike saliency detection,the target of salient object detection is to recognize whether the image contains salient objects and to locate individual objects rapidly and accurately from salient regions of interest in the human vision.Salient object detection in images has important value to the research of object segmentation,object recognition,object tracking and so on.This paper proposes a method of saliency-subitizing computation and salient object detection in images based on the deep learning.First of all,based on the GoogleNet model,we compute the saliency-subitizing of the image,to identify the image as one that contains no salient objects,single or multiple salient objects,which is the prejudgment of salient object detection.Next,according to saliency-subitizing of the image,different strategies are used to do salient object detection.We don't do detection for the images which contain no salient objects,and the images which contain single or multiple salient objects enter the stage of the salient object detection.We improve the YOLO model and propose the multi-layer feature transfer network model.The model training process uses the original images,and the detection process uses hybrid method with original image and saliency map.The method prejudges the image with saliency-subitizing and identifies images that do not contain salient objects,which can avoid blindly detecting salient objects and greatly reduce the computational cost.In the meanwhile,the hybrid method combining the original image and the saliency map implements the accurate location of the salient objects,and solves the problem of distinguishing the aggregate or overlapping objects separately.In this paper,we design and implement the saliency-subitizing computing network contrast experiments,salient object detection network contrast experiments and the contrast experiments between salient object detection without saliency-subitizing calculation and salient object detection based on saliency-subitizing calculation.Salient object detection is performed on the public datasets SOS,MSRA and DUT-O,and the experimental results indicate that our method can effectively improve the performance compared with the state-of-the-art.