Visual Saliency Detection By Simulating Human Vision And Its Applications

Human visual system performance far exceeds current machine vision,the current method of effectively improving the performance of machine vision algorithms is to simulate the human visual mechanism.Human vision has a visual attention mechanism.Through visual attention,the human brain system can use limited resources to quickly handle the most interesting targets in the scene.The model built for visual attention is called the saliency model and is an important research topic in cognitive science,neuroscience and computer vision today.There are two main research directions of fixation prediction and salient object detection in the field of visual attention modeling.The former focuses on predicting human eye gaze in natural images and helps to understand human visual attention.The latter aims to extract prominent target locations and outlines in images and is widely used in high-level task of computer vision.For tasks such as image segmentation,the fixation prediction model performs significantly less well than the salient target detection model.Because the fixation prediction model always produces a sparse point-like area,the significant target detection model produces a smooth target area and contour.How to construct a machine vision algorithm closer to human visual perception,how to make the visual attention model of two research directions learn from each other,complement each other's profit and improve the performance of the algorithm is the goal pursued by researchers.This article focuses on the bottom-up significant target detection algorithm.The purpose of this paper is to simulate the human cranial nervous system with the random right feedforward neural network that can be trained in real time.With reference to human visual mechanism,a new framework of significant object detection system is proposed to obtain a humanvision-aware machine vision algorithm.The main research contents are as follows:(1)Propose a significant target detection framework that simulates human eye gaze.Through the observation and analysis of human gaze,it is considered that repetitive scanning(micro-gaze)and saturation / decay of visual perception are the important behavioral signs of human visual perception.In this way,the initial gaze area is generated by the traditional gaze prediction algorithm,and then the stochastic neural network "online sampling-learning modeling-pixel classification" is parallelized with respect to the gaze area to generate the rough visual saliency map and the preliminary target perception by superimposing the classification results Furthermore,we construct a serial iterative feedback process for the gaze area to make the target perception saturated and generate more detailed saliency maps and targets.A dynamic and positive feedback algorithm framework is formed,which can simultaneously obtain the visual saliency map and the most significant target segmentation result.(2)To simulate the multi-channel characteristics of human visual perception,the gaze prediction part of the above algorithm is improved.Firstly,saliency maps of the original images are obtained by a variety of saliency detection algorithms respectively,and the images are summed to form an integrated saliency map.By binarizing the map,initial salient regions can be formed.Related improvements can significantly improve algorithmic performance.(3)Supervised learning algorithm relies on training samples.When the training sample contains a lot of noise and can not effectively represent the significant target,the algorithm detects the target is often not ideal.In order to solve the above problem,we use the RBD algorithm to suppress the background noise in the saliency map.Firstly,the superpixels of the original image are calculated,the background probability of the superpixels is calculated,and the background pixels of the significant image are suppressed by giving a low weight to the superpixels with high background probability.This effectively reduces the chances of the background pixel being the targetpixel.By making the background suppression of the coarse saliency map in the above algorithm,the system sampling deviation of the machine learning positive feedback iteration in the subsequent algorithm can be greatly reduced,thereby effectively improving the system performance.(4)The above algorithm uses a number of open standard image database experiments,and with the classic,the latest significant detection algorithm for comparison experiments.The experimental results show that the proposed algorithm can obtain the most advanced parameter indexes and the detection of the significant target is closer to the human visual perception.The positive feedback process in the algorithm can be quickly saturated,and does not significantly increase the algorithm burden;suitable as an effective post-processing means,significantly improve the performance of the existing saliency detection algorithm.The final improved algorithm in this paper can be applied to the segmentation of the cell image and the Baidu humanoid library without changing any parameters and structure.The effect is more or less than that of the contrast algorithm,which shows the robustness of the algorithm similar to human vision.