Research On Saliency Detection Algorithm Based On Frequency Domain Analysis And Neural Network

In today’s massive growth of data scale,how to filter effective visual information from the massive image video information and focus visual attention on selecting a subset of visual scenes is especially important to achieve intelligent image processing analysis.Salient object detection is the research direction to study the visual selectivity mechanism and extract visual attention information to filter out redundant interference.After years of research,many classical saliency object detection algorithms have emerged,and the research results have been published in major computer vision conferences and journals,and the relevant algorithms are gradually applied to many fields such as image segmentation,image intelligent cropping,visual tracking,and industrial defect detection.According to different research purposes,salient object detection is generally divided into human eye gaze point prediction and object-level salient object detection.Human eye fixation prediction mainly simulates the process of human eye visual attention and predicts the human eye gaze point region through relevant algorithms,while object-level salient target detection is mainly devoted to extracting the basic contours of salient objects in images or video samples.The current bottom-up algorithm for human eye fixation prediction suffers from the problem of missing detection of salient small objects in images,and the object-level salient object detection also suffers from unclear image edges and inaccurate detection results.In order to improve the detection algorithm,two detection algorithms are proposed in this paper to address the problems related to salient object detection.The main work of this paper is as follows:For the existing human Eye fixation prediction on the lack of small object region extraction,this paper proposes a multi-scale human eye gaze point prediction algorithm based on frequency domain,which can predict the human eye gaze point by simulating the human eye gaze process through multi-scale wavelets and filtering the redundant information in the image.The specific method is: transforming the image to frequency domain,simulating the human eye visual attention mechanism by multi-scale wavelet,and getting the human eye fixation prediction.Compared with the traditional algorithm,the algorithm in this chapter is more accurate in perceiving and locating small objects.The existing object-level saliency object detection algorithms have problems such as poor detection of image edges,incomplete image contours,and inaccurate salient object detection regions,etc.The Transformer-based saliency algorithm proposed in Chapter 4 of this paper crops the input natural images into image blocks of the same size for input model,trains the neural network model through the self-attention mechanism,and in the image In the feature extraction stage,the global feature information is extracted comprehensively,the semantic information of the image is extracted by multi-scale and the saliency object detection area is gradually optimized by multi-stage fine-tuning.The experiments show that the detection algorithm of this paper is more effective and robust for natural image detection of complex scenes,and the detection results are closer to the real scene labeling results.