Detection And Recognition Of Salient Objects Based On Deep Learning

The ability of salient object detection is quickly extract salient regions in an image by mimicking the human visual attention mechanism.It enables the computer accurately and quickly locate the area of interest in the image for processing.However,the saliency detection algorithm based on deep learning has the problem of insufficient utilization of image features.In this paper,the performance of saliency detection improved based on the complementary aspects of high-level and low-level feature information extracted from the network.Apply saliency detection to image recognition:first,obtain a saliency map of the image by detecting saliency objects,and then segment the image by the saliency map to extract the salient area of the image.Finally,image recognition performed on the salient area of the image.This paper focuses on improving the detection performance of salient objects and improving the image recognition rate of convolutional neural networks.The main tasks are:1)Aiming at the problem of insufficient use of high-level and low-level features of the image by the deep guide network,the problem of forged or incomplete saliency regions,and the problem that the traditional convolution module cannot highlight the salient features well.Use topology maps and group convolution modules to improve deep guide networks,and a saliency object detection algorithm based on the improved deep guide network is proposed.After experimental verification,the proposed algorithm uses PR curves,maximum F-Measure,and MAE values as the evaluation criteria on the datasets ECSSD,DUT-OMRON,and HKU-IS.Compared with other algorithms,it has been achieved the best results.2)In the image recognition network,in order to solve the problem of poor performance of Softplus activation function in convolutional neural networks,a PSoftplus activation function proposed to improve its shortcomings.After experimental verification,the recognition rates of convolutional neural networks based on improved Softplus activation functions on MNIST,CIFAR-10,and CIFAR-100 are 98.69%,90.35%,and 67.24%,respectively.Compared with Sofplus and other activation functions,it has achieved the best results3)Applying the saliency detection algorithm proposed in this paper to image preprocessing for image recognition.First,the saliency object detection algorithm based on the improved deep guided network obtains the saliency map of the image,and then the salient part of the image extracted based on the saliency map for image recognition.After experimental verification on convolutional neural networks based on improved Softplus activation functions and other more complex networks,the pre-processed image recognition rate has improved to a certain extent,and the best can reach about 0.82-1.66 percentage.Figure 38 table 6 reference 90...