In computer vision technology, semantic segmentation technology occupies a very important area, which is widely used in driverless and other fields. Semantic segmentation of urban streetscape image is a difficult task, improving segmentation accuracy has been one of the ultimate goal for a long time. There are some problems in segmentation accuracy, including insufficient access to context information and the dim segmentation results at the edge of different objects. Here, based on the full convolution neural network (FCN) in deep learning, we select duel attention network (DANet)1 as our baseline, which introduces attention mechanism to detect context information and its mIoU on Cityscapes reaches 0.646 and pixAcc reaches 0.941. Besides, we try to get richer multiscale context information by replacing the position attention module (PAM) with compact position attention module (CPAM) . In addition, we use a loss function based on distance to edge and the number of new pixels to adjust the imbalance between positive and negative samples. Finally, compared to the baseline, the former figure rises 1.5 percent and the latter rises 1.8 percent. The accuracy of semantic edge segmentation is improved.
As a newly proposed research topic in recent years, panoptic segmentation is the combination of semantic segmentation and instance segmentation. The difficulties of panoptic segmentation include not only the common problems in traditional segmentation tasks, such as small object segmentation and low edge segmentation accuracy, but also the two-way fusion and the determination of conflicts. Considering the defect of edge segmentation, we use the edge optimization module SegFix based on edge detection and direction prediction for edge optimization, which reaches higher accuracy and shortens the calculation time by 3 times compared with Dense CRF. Based on the dual CNN fusion, we select EfficientNet as the baseline, which is more efficient, with depthwise separable convolutions and Mask R-CNN to achieve two-way segmentation. In addition, we use SegFix in multiple panoptic segmentation models to verify its versatility in panoptic segmentation. Finally, our PQ on the Cityscapes validation set reaches 65.5, which achieves the state-of-the-art result with all other panoptic segmentation models under the same experimental conditions, and we confirm that the edge optimization algorithm we use is universal for panoptic segmentation, and its consequence is better than other edge optimization algorithms.
Scene classification for Remote sensing image has attracted great attention because of its difficulties and wide application. There exits several limitations for traditional CNN-based methods, such as insufficient feature extraction ability and complex target of remote sensing image features. In addition, the experimental data is based on the overhead view, which is characterized by fuzzy semantics, small differences between classes and significant differences within classes. To address those issues, we realize several classic network improvement methods such as transfer learning and introduce the attention mechanism Squeeze-and-Excitation (SE) module. We carry out the fine-grained analysis of the space-based view scene image, specifically using the progressive multi-granularity puzzle training for scene recognition. We also propose a semantic-driven scene fine-grained enhancement based on the classic classification network and the progressive multi-granularity puzzle training. To verify the effectiveness of the proposed semantic-driven scene fine-grained enhancement model, we conduct comparative experiments based on several widely used CNN models and a public remote sensing image scene classification data set, and achieve the state-of-the-art result on the data set.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.