Researches On Salient Object Detection Method Based On Visual Saliency

With the rapid development of the society,the processing of massive data for pictures and videos has become increasingly important.Extracting interesting parts from them is an important research topic.The human visual attention mechanism enables humans to quickly search and locate areas of interest in the face of natural scenes.The introduction of the human visual attention mechanism in computer vision tasks,namely visual saliency,can bring great help and improvement to visual information processing.The significant object detection algorithm developed by visual saliency can quickly extract important information from the image and video,which helps to allocate limited computing resources to important information and has important applications in object recognition,image compression,image retrieval,and image redirection.This chapter explores the theoretical basis and research direction of significant object detection,and proposes two saliency object detection algorithms that are adapted to different application scenarios.Firstly,a saliency detection algorithm based on Guided Boosting learning without supervision is proposed to improve the detection effect.This algorithm proposes a new color feature extraction method.Based on this,the SLIC superpixel segmentation is performed on the original image first,and then the super-pixels are calculated based on the bottom-up model to calculate the feature contrast of the super-pixels.Select the sample set,perform the cellular automata optimization on the rough sample set to establish the guide sample set,and finally use the sample seeds to guide the sample selection to supervise a strong classifier for each image based on the Adaboost learning algorithm for each image.The superpixels are used to predict the strong saliency values,Integrate this values and saliency values in the guiding sample set,a saliency map is obtained.The saliency detection algorithm achieves good results in a simple database and application scenario,and based on self-supervision,there is no need to manually mark the original image.Considering that the first algorithm is not effective in the detection of complex scenes,a saliency detection algorithm based on deep learning and multi-tasking network is proposed to detect the foreground target category and the pixel-level label's salience target of the image-level tag.Detection of joint learning,fusion of heuristic significance priors in the network,and the addition of pyramid pooling modules and global smoothing pooling modules,resulting in a significant improvement in detection accuracy and rapid saliency detection algorithms on complex databases Robust effects can also be achieved and can be well adapted in the complex environment of natural scenes.Through experimental verification,the two algorithms proposed in this chapter have great advantages and can adapt to different application scenarios and have certain research significance.