Research And Application Of Image Salient Object Detection Based On Convolutional Neural Network

In today's booming society of big data technology,massive amounts of data need to be quickly streamlined to meet actual needs,and the human vision mechanism provides a paradigm for data processing algorithms.The visual system can quickly find the object of interest in the picture and pass it back to the brain for further processing.The introduction of the visual system's attention mechanism in the computer can greatly improve the performance of the computer.The saliency calculation in the visual attention mechanism can provide fast computing capabilities to locate significant foregrounds,exclude backgrounds,and accurately complete various image analysis tasks.Salient object detection is widely used in a variety of computer vision tasks.Based on the theoretical basis of salient object detection and the problems of current methods,this paper focuses on the salient object detection model based on convolutional neural network and the salient object grading framework combined with image inpainting algorithm.First,this paper proposes a MSC-Net image salient object detection model based on convolutional neural networks for small objects that are detected by existing salient object detection methods.This model uses the multi-scale fusion module,multi-scale pooling module and joint loss function designed in this paper.The multi-scale fusion module provides neurons with convolution kernels of different receptive field sizes,the multiscale pooling module directly transmits shallow layers of information to deep layers and the joint loss function penalizes the wrong pixels with the total variation loss function,and uses the structural similarity index loss function to constrain the boundaries of the image.The experimental results show that the end-to-end network MSC-Net proposed in this paper can achieve salient object detection results with high accuracy and clear edges.At the same time,the comparison of the network with other algorithms on public datasets reveals that the subjective and objective performance of MSC-Net both are better than other salient object detection algorithms.In order to further expand the application of salient object detection technology,this paper proposes a salient object grading detection framework which combines image inpainting algorithm and salient object detection algorithm.The salient object grading can be completed by iterative salient object detection and image inpainting.Among them,the model of the salient object detection part adopts MSC-Net,and the pictures taken in the training set are newly added to approximate the actual use.The image inpainting part improves the multi-loss function based on the deep learning based image repair network,and achieves better repair results.The framework can perform grading detection on salient objects in pictures without a specially labeled training set,and provides an effective solution to the grading problem.