利用卷积神经网络的显著性区域预测方法

针对神经网络的显著性区域预测存在数据采集代价大、处理繁琐等问题，提出2种卷积神经网络，即从头开始训练的浅层卷积神经网络，以及前三层源自另一个网络的深层卷积神经网络。其中，浅层网络结构简单，可避免过拟合问题；深层网络可以充分利用最底层的模型参数，收敛更快，效果更好。所提卷积神经网络应用于回归问题，均没有直接训练特征图的线性模型，而是在迁移层上训练了一堆新的卷积层。从端到端的角度解决显著性预测，将学习过程演化为损失函数的最小化问题。测试和训练在SALICON,SUN和MIT300数据集上进行，实验结果验证了所提方法的有效性。其中，深层网络和浅层网络在SALICON和SUN数据上的结果相似，深层网络在MIT300上的结果更优，与其他方法相比，所提方法具有不错的表现，而且具有跨数据集的鲁棒性。

A significant regional prediction method using convolutional neural network

here exist the problems of high cost of data acquisition and tedious handling in the significant regional prediction based on neural network. Aiming at the problem, two kinds of convolutional neural network are proposed, which are shallow convolutional neural network training from the beginning, and deep convolutional neural network with three layers from another network. Among them, the shallow network structure is simple, which can avoid overfitting problem. Deep network can make full use of the lowest model parameters, and converge faster and make better results. The proposed convolutional neural network is applied to the regression problem, and there is no linear model to train the feature map directly, but a new heap of convolution layers is trained on the migration layer. The saliency prediction is solved from the end to end frame, and the learning process evolves into a minimization problem of loss function. Testing and training are carried out on SALICON, SUN and MIT300 data set. The experimental results verify the effectiveness of the proposed method. Among them, the deep and shallow network in SALICON and SUN data is similar, the deep web in the MIT300 is better. Compared with other methods, the proposed method has good performance and is robust across data sets.