Image Saliency Target Detection Based On Multiple Deep Features

The arrival of the era of big data is accompanied by an exponential growth of data volume.As the most intuitive and convenient way to record and express information,image processing has always been a popular topic for researchers.How to imitate the mechanism of human visual attention and detect the salient object in the image is one of the hotspots of the current computer vision research.In recent years,lots of methods based on deep learning have been proposed.The emergence of convolutional neural network immensely improves the existing saliency detection models.The paper focus on fully mining the deep information of different aspects in the image.The model can get high accuracy through multiple features.In this paper,the main defects of traditional methods and common deep learning methods used in image salient object detection task are listed and analyzed,and the corresponding solutions are proposed.The main contributions of this paper are as follows:In order to solve the problem that it is difficult to determine the salient regions and the multiple salient objects,a network structure of multi-level progressive deep feature aggregation is proposed.This method improves the disadvantages of traditional deep network,such as one-way structure and single feature,and the model uses multi-layer intermediate feature maps to participate in the fusion of the final saliency feature maps.The network structure is divided into three stages.The multi-layer convolutional encoding stage obtains feature maps of different levels,and the upsampling and decoding stage upsamples the feature maps of each level to the appropriate size.Finally,these upsampled feature maps are fused in the fusion stage.This method draws on the idea of multi-scale fusion in traditional methods,and has also achieved good performance in deep learning tasks.The experimental results show that the method has a superior performance in the location of the salient region.On this basis,this paper also explicitly models other deep information of the image.In order to solve the problem of low contrast of the images,this paper proposes a multi-channel feature re-weighting module and a multi-channel feature fusion module based on the channel attention mechanism.Since the salient object detection task is essentially a kind of imitation of human visual attention mechanism,this kind of module structure can fully mine the deep channel feature information of the image,find out the channels that have a greater contribution to the salient stimulus and reweight them,highlighting its position in the feature map.Firstly,the image is coarse coded by the global pooling layer in the convolution stage,and then the vector relearns the salient stimulus contribution of the feature channel through the adaptive reweighting part.Finally,the reweighted vector is multiplied and accumulated with the original map to form a new feature map.Experiments show that this method can significantly improve the performance of the detection model on several challenging public datasets.For the problem that the contours of salient targets are blurred,this paper proposes the salient object contours auxiliary detection task.By introducing branches,the contour information of the target is explicitly modeled.The branch structure of this task is similar to the main structure.Since the task aims to detect the outline of the salient object,the number of network levels is slightly reduced compared to the main structure The branch makes the final detection network become a multi-task neural network structure with the contour detection as the auxiliary task.Experiments indicate that the auxiliary detection can clearly refine the boundary information,make the target contour clearer,and greatly improve the quantitative indicators.Finally,this paper fuses the above three modeling methods of deep feature to form a multi-level,multi-channel,multi-task feature fusion salient object detection network,and briefly describes the fusion methods of each deep feature.Through qualitative and quantitative comparison with other methods,the advantages of this fusion method are proved.In addition,by adding ablation experiments,the effectiveness of the proposed network structure components and the application prospect are proved.