Research And Design Of Image Saliency Detection Algorithm Based On Convolutional Neural Network

Image saliency detection refers to the use of the attention mechanism of the human eye to detect more prominent target areas in the image.When facing a natural image,humans can accurately capture salient areas of the image.For example,when the background is a blue sky,the white clouds dotted in it are prominent.Another example is a ship on the lake,the ship is also significant relative to the lake.These scenes have common characteristics.The contrast between the salient part and the background part is large,and the color contrast is also large.For scenes with a single background as mentioned above,there are many simple and highly accurate image saliency detection algorithms.However,for scenes where the background of the image is more complex and the contrast between the salient target and the background is small,the existing algorithms do not perform well for this type of image.Therefore,a more robust saliency detection model needs to be constructed.This paper takes VGG16 as the basic network and builds an end-to-end saliency detection model.First,a significant prior is added to the input of the network;then,this significant prior is input into the convolutional neural network for training with the original image;finally,the final model parameters are determined by adjusting the pre-training parameters of VGG16.For the problems of complex image backgrounds,the last layer of each set of convolutions of the VGG16 network is stitched,and the shallow output image with edge texture details and the deep output image with semantic level information are fused to enhance the network learning ability.Aiming at the problem that the contrast between foreground and background makes the salient map difficult to accurately distinguish between salient subjects and inconsistent edges,this paper proposes to enhance the original image with contrast enhancement,sharpen the edge information,and then add the result set to the training set.Through this process,the neural network can learn more edge information and enhance the robustness of the algorithm.The algorithm proposed in this paper is evaluated on three public saliency datasets.Experimental results show that the algorithm can effectively highlight the saliency area.On the SOC data set,it can be seen that for images with small foreground and background contrast and cluttered background,the algorithm in this paper can accurately find significant parts and generate coherent edges.Compared with other algorithms,F-measure and MAE also have greater advantages.