Salient Object Detection Based On Visual Attention Mechanism

Salient object detection aims to detect the most prominent object in the image.The traditional salient object detection method can obtain ideal results in simple scenes.However,in complex scenes,the detection results of traditional methods cannot reach a satisfactory level.Therefore,this paper proposes two salient object detection models based on convolution neural network,which combine visual attention mechanism to achieve high-performance or efficient salient target detection.(1)Attention-guided Context Feature Fusion Network(ACFFNet).Its detection model consists of an encoder and a decoder.The encoder is Res Net-50.The decoder is composed of the multi-field channel attention module,the context feature fusion module and the self-refinement module proposed in this paper.In addition,the cross-consistency enhancement loss is proposed to guide the network learning(2)Position Prior Attention Network(PPANet).The decoder of the network includes the residual optimization module,the atrous spatial pyramid pooling module,the position prior attention module and the feature integration module.Encoders can choose Mobile Net or Res Net-50 to meet the application requirements of different scenarios.In addition,this chapter proposes the boundary enhancement loss to mitigate the adverse effects caused by the scale changes in the saliency map and boundary map.The experimental results on five popular benchmark data sets show that the proposed PPANet-R is superior to the existing saliency object detection model,and the lightweight PPANet achieves the same precision as the most advanced heavyweight saliency object detection method at the real-time speed of 150 FPS.The ACFFNet and PPANet proposed in this paper are verified on five public data sets,and the results prove their effectiveness and superiority.