Design And Application Of Saliency Detection Algorithm With Multi-module Fusion Based On Deep Learning

Saliency object detection is one of the research contents in the field of computer vision,which is also called saliency detection,by computer simulating the human eye to focus on the most attractive region in the scene as the local saliency region.As a preprocessing technique for images,saliency detection has a wide range of applications in several scenarios,such as visual tracking,object recognition,and image segmentation.At present,the better algorithms are mainly network models built with neural networks as their core,and their performance has been greatly improved compared with traditional methods.However,the problem of effective fusion of multi-scale features and loss of high-level semantic information remains a major challenge in this field.In addition,many algorithms ignore the overall model computation in order to improve the detection accuracy,which makes the algorithms extremely limited in application,so how to balance the algorithm performance and computation is also an issue of concern.To address the problems described above,this paper conducts a related exploration,and the main research work is as follows.Aiming at the problem of how to extract and fuse multi-scale features,this paper proposes a multi-scale balanced attention interaction saliency detection algorithm.This algorithm adopts two processing mechanisms for features of different depths.For advanced feature extraction,this paper proposes a balanced attention model.Based on the calculation of the affinity between the features within the sample,the samples can interact with each other,and the two work together to capture the advanced features.For feature extraction at other levels,this paper proposes an interactive residual network model that adaptively extracts features at different scales by using the information of two resolutions to perform two crossover operations with each other.In order to effectively fuse all the features,this paper introduces a bidirectional propagation network model,which is added step by step in the direction from deep to shallow and from shallow to deep,and the output result is used as the final prediction result.Experiments show that the algorithm has good performance,especially on the DUT-OMROM dataset,the F-measure reaches 0.763,and the MAE reaches0.069.To address the problems of feature loss and model lightweighting,this paper proposes an asynchronous cascading saliency detection algorithm based on hybrid attention.The algorithm does light weighting on the basis of the self-attentive model,pools the twodimensional features before classification to reduce the exponential computation,and combines the balanced attention feature to prevent feature loss and capture contextual information jointly in space and channel.In order to increase the characterization ability of shallow features and suppress noise,this paper improves on the two-way propagation and proposes an asynchronous cascade strategy,which can skillfully integrate the outputs of different models effectively by means of feedback and use multi-layer supervision for prediction.Experiments show that the algorithm can achieve better results on standard datasets,especially MAE,which achieves 0.085 and 0.039 on PASCAL-S and HKU-IS datasets,respectively.