关键词: 显著性检测, 深度神经网络, 生成对抗网络, 小型生成器网络, 融合算法

Abstract: Aiming at the difficulties of model training,large amount of model parameters,and slow detection speed in the saliency detection method based on deep neural network models,this paper proposes a saliency detection method that integrates small deep generative models.The method uses the generative adversarial network as the framework,designs a discriminator network consisting of 11 convolution modules and 5 pool layers,as well as a small generator network that does not contain the pool layer,including only 15 convolution modules and 5 transposed convolution modules.Among them,the size of the small generator network is only 2.4M,and the amount of parameters is only about 670,000.A trained small generator is used for saliency detection and fused with the LMB(Salient object detection based on background block re-selection method) algorithm through the designed fusion algorithm to get the final result.A large number of experiments and comparative analysis show that the proposed method has achieved significant improvements in both F value and MAE (Mean Absolute Error) value.

Key words: saliency detection, deep neural network, generative adversarial networks, small generation network, fusion algorithm