Research On Saliency Detection Method Based On Depth And Width Neural Network

Image saliency detection is to segment the most important areas in an image.Often the problem of image saliency involves computer vision,neuroscience,cognitive psychology and other fields.In recent years,with the great achievements of deep learning in the field of computer vision,the application of deep learning has also played a good role in the problem of image saliency detection,so the algorithm based on deep convolutional neural network has become most effective solution for solving image saliency problem.But as far as the current saliency algorithms based on deep convolutional neural networks are concerned,most of them employ very complex deep convolutional network models for saliency detection,which will rely on high computing equipment to complete the implementation of the algorithm,making these algorithms impossible applied to mobile devices,nor effectively complete the preprocessing process of other advanced visual problems.For researchers,improving the computational efficiency of neural network based saliency detection algorithms generally starts from two perspectives.One is to tailor the network structure and process it in combination with traditional feature extraction methods.The other is to use a more lightweight network to solve the saliency detection problem.According to these two ideas,this paper proposes two efficient and accurate saliency detection algorithms based on neural networks.Firstly,a deep convolution saliency detection algorithm that combines the lowdimensional feature extraction layer is proposed.In this thesis,an low-dimensional feature extraction layer is proposed to perform feature extraction and map it to the straightened feature maps which are then used as inputs for the full convolutional network.The purpose is to reduce the dimensionality of the data to reduce the amount of calculation.Then,the network joins the full convolutional network for end-to-end learning saliency maps to enhance its low-dimensional feature learning ability.Then,this thesis also uses three loss functions in the training process to ensure the stable convergence of the algorithm and enhance the output effect of the algorithm.Secondly,some recent studies have shown that broad neural network combined with lift learning can achieve excellent performance on classification and regression problems,and this method consumes less time in model training and has high computational efficiency.Based on the above observations,this paper attempts to take advantage of the broad neural network and combine it with lifting learning to implement a saliency detection algorithm model.In this algorithm,the training of this model does not use the entire image as input,but manually extracts regional feature descriptors from the image as input to let the network model learn the feature,and then predicts the significance value for each regional feature.In addition,the algorithm also constructs a random field with parameter learning conditions on the saliency map to refine the prediction results,and has obtained more accurate results.Finally,for each algorithm proposed in this paper,a wealth of experiments have been performed to verify the innovation of the algorithm.The saliency detection algorithm in this paper will be compared with the current eight representative saliency algorithms on four public data sets to test and compare PR curves,-measure,MAE,runtime,and visual effects.Experiments show that the algorithm has better detection accuracy and faster running speed.