Image Saliency Detection Based On Visual Attention Mechanism

The saliency detection is mainly based on the interdisciplinary subjects such as psychology and biological neuroscience,and simulates human visual system to predicate the interest area of the image,which can provide effective preprocessing results for image segmentation,image retrieval and other image processing fields,and improve the processing efficiency of subsequent tasks.In view of that problem that the current algorithm lack sufficient biological theoretical basis,the limitation of artificially define features and the insufficient accuracy of the saliency detection,two algorithms in different views of the saliency detection are proposed.Because of the problem of the way of the calculating is single and the distinct area is blur in the algorithms,this paper analyzes its detection mechanism,improves its characteristics,and proposes a division normalization algorithm based on the SLIC(simple linear iterative-clustering)and spatial weighting.The algorithm is added to SLIC,which replaces the pixels by the areas,and then builds spatial weights based on the Gestalt principle to further emphasize the distinct regions of the image,and then optimize the output to get the final results of detection.The existing salient model only increases the number of significant features without deepening the analysis and utilizing features well,and most of the salient model applications are small in scope and cannot satisfy the detection in complex situations.In this paper,a significant detection algorithm based on deep convolutional neural network and regional contrast is proposed based on the traditional saliency detection principle and the technique of deep learning.The model is based on the preliminary significance of network prediction based on VGG-16 network,and the regional contrast method is added to increase the difference between the significant target and the background region and improve the significance detection effect.In this paper,the two algorithms and contrast algorithms are evaluated by the database MSRA-1000,and the algorithm is evaluated by both subjective and objective measures.Experimental results show that the salient image of the two algorithms can generate the more accurate salient areas,which are more consistent with human vision,from the objective evaluation index,such as Precision-Recall curves,ROC curves and Fmeasures and MAE are higher than other contrast algorithm,which can be seen that the comprehensive performance of our algorithms are more outstanding.