Research On Saliency Object Detection Algorithm Based On Deep Learning

With the in-depth study of deep learning and the continuous development of big data information technology,the application of artificial intelligence is becoming more and more mature,and people’s life has entered the era of intelligence and information.As an important branch of artificial intelligence,deep learning-oriented computer vision technology has made more remarkable achievements in this era,breaking through the bottleneck of traditional algorithms.Salient object detection,as an important research direction in the field of computer vision,has attracted more and more attention,and has been widely used in artificial intelligence,intelligent medicine and other fields.As a preprocessing method for image segmentation,image detection and recognition,it can effectively save the time cost of image processing.This thesis studies the significance detection algorithm based on deep learning,mainly introduces the principle of the current mainstream significance object detection algorithm,compares various algorithms,and summarizes and concludes the existing significance object detection algorithms.Aiming at the problems of large parameter number and complex image background in the domain of salient object detection,this thesis proposes a salient object detection model based on Pool Net algorithm.In order to reduce the number of model parameters,Dense Net network with narrower network and fewer parameters is used as the backbone network,thus reducing the computing overhead of the backbone network.In order to enhance the feature extraction and improve the detection performance of the model,a new bidirectional weighted feature pyramid network(Bi FPN)was used to improve the accuracy of the model.In order to solve the problem of unclear boundary of significant objects,the mixed loss function obtained by weighted summation of binary cross entropy(BCE),structural similarity(SSIM)and other loss functions is applied to the improved model.The results show that compared with the original algorithm,the accuracy rate of the improved Pool Net algorithm is increased by 1.02%,the number of model parameters is reduced by 80%,and the definition of significant object boundary is improved,which has good robustness.There are more and more algorithms for significance detection of RGB-D images based on deep learning,but how to effectively use Depth feature information to assist significance detection is still a difficult problem.Since different levels contain different feature information,it is also a very important research content to study how to integrate feature information of different scales.In order to solve these problems,an improved CDINet algorithm is proposed to detect the saliency of RGB-D images.Firstly,a multi-scale feature fusion module is added to enhance the transmission of feature information between the encoder and decoder,effectively reducing the shallow features lost in the process of network convolution,and obtaining more feature information of salient objects through the jump connection of auxiliary decoder.Secondly,a circular attention module is connected at the tail of the network structure of CDINet.By using the memory-oriented scene understanding function,the local details are gradually optimized to further improve the detection performance of the model.Finally,the loss function is adjusted and the consistency enhanced loss(CEL)processing is used.Because of the spatial consistency and other problems caused by the fusion of features of different scales,the salient areas can be uniformly highlighted without increasing parameters.The improved model was tested on LFSD data set and STERE data set and the results were analyzed.Studies show that F-measure increases 0.6% and 0.4% respectively,and MAE decreases 0.4% and 0.3%respectively.Compared with other algorithms,the model detection performance is better and the adaptability is higher.