| Image saliency detection refers to the technology of using computer to simulate human visual attention mechanism.The main task is to quickly and accurately detect the areas most concerned by human eyes in an image by establishing an image saliency detection model or algorithm.As an important step of image preprocessing,image saliency detection plays an important and irreplaceable role in many traditional image processing tasks.In order to clarify the advantages and disadvantages of traditional algorithm and Deep-learning algorithms,this thesis summarizes the research status of saliency detection at home and abroad,and then carries out the following work.(1)Because the FT algorithm is not ideal when the background environment of salient area is complex and the salient area is large.According to the characteristics that people are more likely to pay attention to the image central object,this thesis based on FT algorithm to enhance the central area,normalize the LAB color eigenvalues,and weighted the threechannel information of LAB.Through the analysis of the advantages and disadvantages of the traditional saliency detection model,we find that although the traditional algorithm has advantages in detection speed,it has great defects in detection effect.Traditional algorithms can not meet the requirements of saliency detection,such as emphasizing the largest salient object,uniformly highlight whole salient regions,establishing well-defined boundaries of salient objects,risregard high frequencies arising from texture and noise,and blocking artifacts.(2)By analyzing the characteristics of some existing saliency detection algorithms based on deep learning,we find that the existing depth model has the following problems.Firstly,existing Deep-learning models need to pool feature maps after convolution.Because the feature map becomes smaller gradually in the convolution process,the size of the final detection result is not consistent with the size of the original input image.At the same time,some existing saliency detection algorithms based on deep learning method use fullconnected network in the last part of the network model,which makes the spatial information of feature map not fully used,resulting in the waste of spatial information.Furthermore,these models mainly extract high-level features from the convolution layer at the end of the network and combine them in a non-linear method,but the effect is still not ideal due to the lack of low-level visual information such as object edges.In order to solve the problems of existing models,this thesis proposes using SEGNET as the basic network to realize saliency model detection.Through full convolutional coding and corresponding decoding,the model perfectly overcomes the inconsistency between saliency map and original input image and the incomplete use of spatial information of feature map.By constructing a symmetrical network,the model solves the problems of depth information loss and lack of effective use of low-level visual information such as object edges in the process of pooling.By modifying SEGNET image segmentation network,this thesis transforms the original image segmentation network based on cross-entropy classification into a saliency detection model based on regression algorithm.Next,this thesis carries out migration learning on the redesigned network framework on THUS and HKU-S datasets and carries out comparative experiments.By comparing with other saliency models,we can see that our improved algorithm has considerable advantages over other algorithms and basically solves the problems mentioned above. |
