A Saliency Detection Model Based On Bifurcated Convolutional Neural Network With Attention Application

The saliency detection task is a task that uses computer vision technology to simulate the Attention distribution of the human eye to different regions of the image.The model developed in this task is widely used in other visual tasks such as object segmentation,object proposal generation,image cropping,etc.,becoming an effective image preprocessing method.In the existing saliency detection research,U-shape structure is widely used.These structures have been proven to effectively achieve high accuracy prediction in image segmentation tasks,and the model parameters that can be reused for object detection tasks with large amounts of data make their training time and accuracy reach the ideal level.Many researchers have introduced the attention mechanism into the saliency detection field to enhance the ability to extract key area information and make it play a role in the decoding process of image information layer by layer,which effectively improves the utilization efficiency of image coding information but also makes the model the computational burden of the decoding process is greatly increasedIn order to solve the problem of excessive computational consumption caused by the multi-level application of attention mechanisms in the network,we adopted the method of using the salient object prediction as the distribution of attention weights to enhance the model's ability to distinguish between foreground and background.Taking the bifurcated network as the backbone of the network,by designing the attention mechanism to establish the connection between the detection results of the two branches of the network,we can obtain the model's greater perception of the foreground with less parameter cost.In addition,in order to solve the problem that the edge of the foreground area is fuzzy in the prediction of past saliency detection models,we embedded edge weights in the attention mechanism module to enhance the model's ability to obtain information in the edge area.From the actual detection effect,it can be found that the attention mechanism structure we designed optimizes the internal continuity of the foreground area,can obtain a clearer boundary between the foreground and the background,and benefits from the reduction in the amount of parameters to achieve 120 fps detection speed.