Research On Salient Object Detection Algorithm Based On Level Set And Attention Mechanism

Salient object detection is a basic problem in image and vision research.Salient object detection is widely used in computer vision tasks such as image compression and object recognition.Salient object detection is often the superior task of computer vision task,so the accuracy and efficiency of salient object detection are particularly important.The confidence and recall ratio of traditional models are low.Although the deep learning methods can locate the object better than the traditional algorithms,the boundary is usually not detailed enough,and the detection accuracy is insufficient.In addition,if too many convolution layers are added to shallow features,the model will be heavy,and the noise of low-dimensional features will easily pollute high-dimensional features,which will reduce the recognition effect of the model on multiple targets.Therefore,it is particularly important to obtain semantic information by efficiently processing the features of global features and local information.Based on the above background and problems,this paper focuses on the boundary ambiguity of salient object detection algorithm,how to efficiently obtain semantic information and improve multi-object recognition.To solve the problems of fuzzy boundary and low confidence rate,this paper proposes a significance object detection algorithm(RF2Net),which combines the traditional level set method and deep learning.RF2 Net incorporates the idea of loss of level set structure and the reverse attention mechanism.First,RF2 Net uses a new loss function that combines BCE losses,weight level set losses,and weight mean absolute error(MAE)losses.With the level set loss operator,the whole image can be better focused,rather than pixel-by-pixel supervision like BCE loss.The introduction of reverse attention mechanism can effectively reduce the noise in the process of interlayer feature fusion and achieve the purpose of improving the accuracy.Experimental comparisons with 12 stateof-the-art algorithms on 4 datasets show that MAE,max F and avg F algorithms are superior to other algorithms on HKU-IS datasets.Meanwhile,ablation experiments were performed on DUTS and ECSSD datasets to verify the effectiveness of the algorithm.The ablation experiment results show that this algorithm can effectively improve the effect of salient object detection.In order to deal with semantic information and multiple objects effectively,a salient object detection algorithm(DASL)is proposed in this paper.In DASL,a level set attention mechanism module is proposed to process the features obtained from the trunk and obtain richer semantic information to better process the foreground and background for the subsequent network.To solve the multi-objective problem,it proposed a structured loss.Using BCE alone is pixel-by-pixel and does not focus well on the whole.The experimental results show that the proposed loss can output clearer significance graph,and the proposed loss function is much better than the BCE loss.The DASL algorithm showed the most advanced performance when compared with other SOTA algorithms on four popular salient object detection datasets.