Research On The Method Of Image Salient Object Detection Based On Convolutional Neural Network

Salient object detection simulates human visual features by intelligent algorithm to extract salient regions and targets in images.In recent years,salient object detection has become an increasingly critical task in computer vision and its related fields.It can be used as a pre-processing step for visual tasks such as target tracking and image compression,etc.An ideal salient object detection algorithm has higher computational efficiency and accuracy.This paper designs a network which can thoroughly learn effective features.The model improves computational efficiency while maintaining accuracy.The main research work is as follows:First of all,for the detection efficiency problem,we modify the image segmentation network of the ERFNet and get the efficient salient object detection model.Firstly,the residual decomposition structure is used in the network,and the parameters are reduced by decomposing the two-dimensional convolution kernel into a one-dimensional convolution kernel,which can achieve the purpose of improving network detection efficiency while maintaining the original accuracy.Secondly,the encoder-decoder structure used in the model has been improved,and the network layers are deepened to ensure the detection accuracy.Finally,the network is trained and the model is validated on the standard dataset,the experiment shows the validity of the novel model.Second,for the balance between detection efficiency and accuracy,this paper proposes a salient object detection model based on the Balance of Speed and Accuracy Network(BSANet).Firstly,considering the efficiency of model detection,the convolution decomposition structure in the self-driving scene segmentation network is used.Then,according to the characteristics of salient object detection,a special cross-layer connection structure is designed to make full use of the context information.Meanwhile,the multi-scale fusion operation is used to fuse the features that learned by the network at different scales,to achieve the combination of shallow information and deep information required for significant target detection.Finally,through the training experiment of BSANet network,the validity is verified.Finally,in order to objectively evaluate the proposed model,it is compared with nine other advanced models under different evaluation indicators.