| Region segmentation,a branch task of computer vision,aims at detecting certain region and is widely used in autonomous driving,map element detection and video understanding.Compared with traditional region segmentation methods,deep learning based region segmentation methods have shown higher accuracy in public datasets.Region segmentation methods that only rely on high-level semantic features are usually limited by the scale of the target and the results often have unclear boundaries.Utilizing prior knowledge to enhance highlevel semantic features according to specific application scenarios is the key to improve segmentation accuracy.The first part of this paper proposes a road segmentation algorithm which is based on iterative cross-task feature enhancement to solve the incomplete result in remote sensing images.In order to explicitly utilize the correlation between the road segmentation task and the road orientation prediction task to construct the interactive information flow and make full use of prior knowledge in feature enhancement to improve the accuracy of road segmentation results,this paper proposes an iterative cross-task feature enhancement network which jointly optimizes the deep features of the two tasks in a multi-task learning manner.In order to achieve iterative optimization between the two tasks,this paper first designs a semantic-guided feature enhancement(SGFE)module,which uses semantic information to guide and enhance the orientation features to improve the credibility of orientation features,and then designs an orientation-aware feature aggregation(OAFA)module.This module adaptively adjusts the receptive field of the convolution kernel according to the road orientation and improves the direction perception ability of the network to aggregate more effective information.Finally,an iterative cross-task feature enhancement strategy is designed to iteratively enhance the features of two task,solving the problem of insufficient utilization of orientation information.Experiments on public datasets show that the algorithm proposed in this paper has modified the discontinuity of road segmentation results in remote sensing images.The second part of this paper proposes a video salient object segmentation algorithm based on feature enhancement to solve the incomplete and instability results.Compared to salient object detection task,video salient object segmentation task needs temporal information.Optical flow is widely used to provide the temporal information in previous works,which suffer from the computational cost.This paper proposes a temporal memory module based on the attention mechanism to integrate temporal information.The long-range temporal information of reference frames in the video sequence is used to reconstruct and enhance the current frame features.Besides,this paper proposes a temporal consistency loss,which uses optical flow to provide motion information and imposes consistency constraints on the segmentation results to improve the stability.Instead of using optical flow as input,this paper utilizes optical flow to construct the loss function to constrain the segmentation results.Motion information is implicitly embedded into the features during the training process.Thus,no extra computation cost is introduced in inference stage.Compact model can be employed for real-time segmentation.In this paper,the performance of the algorithm is verified on two public video salient object segmentation datasets,and the results show outstanding performance in terms of both accuracy and stability. |
