Research On Weakly Supervised Salient Object Detection Via Noise Robustness

Salient object detection is an extremely important task in computer vision,focusing on identifying and segmenting the most discriminative objects or regions in an image.In recent years,weakly supervised salient object detection has been developed,for reducing the amount of expensive human effort required during ground truth annotation.Weakly supervised methods use pseudo labels converted from weak labels to train the saliency network.However,the converted pseudo labels always contain a large amount of mislabeled noise information compared to ground truth,the noise information in the pseudo labels inevitably propagate to the final predictions.To mitigate this problem,the thesis proposes a noise-robust generative adversarial network and a multi-scale context feature enhancement network.The former makes the saliency network robust to noise information,the latter improves the performance of the saliency network in challenging scenes.Concretely,the innovations of the thesis are as follows:(1)Existing methods are affected by pseudo label noise to generate error-prone predictions.To solve the problem,the thesis proposes a noise-robust salient object detection method based on adversarial learning.The method consists of two parts: noise-robust generative adversarial network and noise-sensitive training strategy.The noise-robust generative adversarial network can separate the saliency information from the noise information from the high-order structure,so that the network is robust to the noise information in pseudo labels.The noise-sensitive training strategy divides pseudo labels into superior or inferior samples according to the amount of noise in each training iteration,so that the network can further balance the learning of saliency information and the robustness of noise information.Compared with other methods,this method not only learn high-order structure information in saliency dataset,but also has sufficient robustness to noise information in pseudo labels.(2)Existing methods are difficult to accurately detect the salient objects under challenging scenes.To solve the problem,the thesis proposes a hierarchical camouflaged data augment method,and designs a multi-scale contextual feature enhancement network to improve the performance in challenging scenes.The hierarchical enhancement camouflaged data augment method divides and selects challenging scenes to complete saliency training dataset.These scenes increase the proportion of challenging scenes in the saliency training dataset,and the network has a deeper semantic understanding of the challenging scenes.The multi-scale contextual feature enhancement network uses a variety of convolution layers to extract and fuse the features of each layer in an all-round way,so that the network still has sufficient information extraction and fusion capabilities for challenging scenes.Experiments show that this method not only keeps the network highly robust to noise in difficult scenes,but also improves the performance of the network for difficult scenes.