An Improved CNN Model for Fast Salient Object Detection

Abstract

In an image, how to quickly and effectively extract the useful regions named target regions in the scene according to the saliency features such as spatial domain, frequency domain etc. for further analysis of salient object detection is one of the challenging topics in the field of image segmentation. Most of the existing salient target detection methods use convolution network to extract high-order semantic features, combine pyramid pooling model to fuse high-order and low-order semantic features, and use Adam or SGD optimizer to optimize the model to obtain the salient object. However, the traditional convolution network model is not optimized for the model parameters, and finally redundant parameters will appear in the model, which will aggravate the training time and practical application detection time of the model. Although SGD is fast, it will fall into a large number of local suboptimal solutions or saddle points in the process of non-convex error function optimization. Adam has better performance, but the speed is slightly slower then t -> ∞ that will not have a good generalization performance. In order to solve the above problems, a new optimization strategy is proposed to compress the model. At the same time, AdaX, an optimizer with SGD speed and Adam performance, is used to optimize the model. Through the test on the open data set DUTS, ESSCD and etc., the proposed optimization model method reduces the parameters of the original model, and also improves the training speed and application detection speed of the model.