Research On Salient Object Detection Based On Neural Network Model Optimization

Salient object detection is one of the most challenging research topics in the field of computer vision.Its purpose is to detect the object area that can most attract human visual attention from the image,so as to improve the efficiency of computer processing of the image.Although some research results have been achieved in this field,the accuracy,speed,scale,and memory footprint of salient object detection models have limited their application in special scenarios or on devices with limited resources.This dissertation takes salient object detection as the research object,and conducts research from two aspects of model structure optimization and model scale optimization,and constructs salient object detection models that meet different needs.The main work of the dissertation is as follows:Aiming at the problem of low accuracy due to the single information propagation direction in the neural network-based salient object detection model,a salient object detection method based on feature propagation is proposed.The direction of information propagation in the classical neural network model structure is from the bottom layer to the top layer,which is prone to the phenomenon of information loss,information isolation and resolution reduction,which seriously affects the accuracy of the model.In this dissertation,through the analysis of the salient object detection method of one-way feature propagation,a salient object detection model based on two-way feature propagation is proposed.The model consists of a forward connection subnet and a reverse connection subnet,inter-jump connections,adaptive learning fusion strategy,and reversal of efficient region-jump connection operations.Experimental results show that this method enhances the flow of information between network layers and effectively improves the accuracy of salient object detection.Aiming at the problem of high computational complexity and slow detection speed due to the large parameter space of the neural network-based salient object detection model,a salient object detection method based on model fusion was proposed.Models based on neural networks have the characteristics of high detection accuracy but require a lot of computing resources and low operating efficiency,while models based on nonneural networks can overcome these shortcomings but have low detection accuracy.This dissertation analyzes the salient object detection method of single model,and proposes a salient object detection model based on dual model fusion.The model consists of preprocessing module,regional flow module,pixel flow module,fusion and post-processing module.The manifold sorting algorithm of the energy equation simplifies the neural network based on two-way feature propagation,so that the non-neural network-based model and the neural network-based model are integrated to achieve the effect of complementary advantages.The experimental results show that the method can effectively improve the speed of salient object detection within a small accuracy variation range.Aiming at the problems of large scale of neural network-based salient object detection models and low detection accuracy of tiny salient objects,a method for tiny salient objects detection based on model pruning was proposed.In the actual operating environment,limited by objective factors such as the location of the acquisition device,the direction of movement of the object,and the shooting conditions,the salient objects in some of the collected images are small.However,the existing large-scale neural network models are interfered by the large background to detect small salient objects,resulting in low detection accuracy.In this dissertation,through the theoretical analysis of the problem of small salient object detection,the Salient Energy Level(SEL)is defined to evaluate the ability of neural network model parameters to distinguish background features and small salient object features;With the help of model pruning technology,model parameters with lower salient energy levels are removed,and a small salient object detection model based on salient energy levels is proposed.The experimental results show that the method can effectively improve the detection accuracy of small salient objects and reduce the scale of the neural network model.Aiming at the large scale of neural network-based salient object detection models and the low accuracy of multiple salient object detection,a multiple salient object detection method based on model pruning was proposed.When there are multiple salient objects in the detection image,identifying the most salient object can effectively improve the model detection efficiency.However,existing methods cannot distinguish the differences between multiple salient objects,resulting in low detection accuracy.In this dissertation,through the theoretical analysis of the multiple salient object detection problem,the Salient Priority Criterion(SPC)is defined to evaluate the ability of the neural network model parameters to distinguish the feature differences of multiple salient objects;With the help of model pruning technology,model parameters with higher saliency priority are reserved,and a multi-saliency object detection model based on saliency priority is proposed.The experimental results show that the method effectively improves the detection accuracy of multiple salient objects and reduces the scale of the neural network model.