Research And Implementation Of Image Saliency Detection By Combining Attention Mechanism And U-Net

With the advent of the era of big data,how to get useful information from massive images has become an important research area in the field of computer vision.Image saliency detection can quickly locate the image area of interest from the complex visual scene by simulating human visual attention mechanism.Nowadays,image saliency detection algorithm has been widely used in image compression,image retrieval,image segmentation and other fields.At present,image saliency detection algorithm can be divided into bottom-up image saliency detection algorithm based on low-level visual features and top-down image saliency detection algorithm based on deep learning.The existing bottom-up image saliency detection algorithms based on low-level visual features can be divided into spatial feature-based image saliency detection algorithm and frequency-domain featurebased image saliency detection algorithm.Based on the experimental analysis,this paper summarizes that the image saliency detection algorithm based on spatial features is suitable for detecting the image with strong contrast of color features between significant target and background,while the image saliency detection algorithm based on frequency features is suitable for detecting the image with weak contrast of color features between significant target and background.At present,there is no algorithm which can detect the two types of images at the same time.In addition,in the field of top-down image saliency detection algorithm based on deep learning,because the model based on U-Net obtains significant information by extracting the high-level features and global information of the image,the extracted significant information lacks the low-level visual information,and the detected saliency map is sparse,irregular and the image boundary is not accurate.In view of the above problems,the main contents of this paper are as follows:(1)In this paper,a new adaptive image saliency detection algorithm is proposed,which fuses spatial and frequency features of image.Specifically,through Res Net18,the algorithm classifies the image into two types: the image with strong contrast between the color features of the significant target and the background and the image with weak contrast between the color features of the significant target and the background.Through the representative image saliency detection algorithm based on the spatial and frequency features of the image,the prospect saliency map of the above two types of images is calculated respectively.Then,it is fused by Bayesian fusion with the background prior saliency map of the input image to obtain the intermediate saliency map,and the conditional random field optimization is used to get the final saliency map.Experimental results show that the detection effect of this algorithm is better than most of the current bottom-up image saliency detection algorithms.(2)An image saliency detection algorithm based on attention mechanism and U-Net is proposed.Specifically,in the U-Net model decoder stage,the algorithm combines the saliency map detected by the above algorithm with the feature image to build multi-scale features fusion to enrich the spatial and local information of the saliency map,uses channel attention mechanism to extract key features to optimize the sparsity and irregularity of saliency map,and constructs the multi-scale weighted loss function to promote the saliency map closer to the ground truth map.The experimental results show that this model has better image saliency detection effect than U-Net model,and it is also better than other mainstream image saliency detection algorithms based on deep learning in recent years.