Bo Zhao

Recent works on semantic segmentation witness significant performance improvement by utilizing global contextual information. In this paper, an efficient multi-granularity based semantic segmentation network (MGSeg) is proposed for real-time semantic segmentation, by modeling the latent relevance between multi-scale geometric details and high-level semantics for fine granularity segmentation. In particular, a light-weight backbone ResNet-18 is first adopted to produce the hierarchical features. Hybrid Attention Feature Aggregation (HAFA) is designed to filter the noisy spatial details of features, acquire the scale-invariance representation, and alleviate the gradient vanishing problem of the early-stage feature learning. After aggregating the learned features, Fine Granularity Refinement (FGR) module is employed to explicitly model the relationship between the multi-level features and categories, generating proper weights for fusion. More importantly, to meet the real-time processing, a series of light-weight strategies and simplified structures are applied to accelerate the efficiency, including light-weight backbone, channel compression, narrow neck structure, and so on. Extensive experiments conducted on benchmark datasets Cityscapes and CamVid demonstrate that the proposed method achieves the state-of-the-art performance, 77.8%@50fps and 72.7%@127fps on Cityscapes and CamVid datasets, respectively, having the capability for real-time applications.

Existing image-to-image translation methods usually adopt an encoder-decoder structure to generate images. The encoder extracts the features of input images using a sequence of convolution layers until a bottleneck, and then, the intermediate features are decoded to the target image. However, the existence of bottleneck layer in such structure may lead to blurry and bad quality of the translated images, since different domain translations may be related to the global or local region in the input image or even in an abstract level. To prevent these problems, we propose the channel attention networks for image translation in this paper. It is a novel model that supports the multi-domain image-to-image translation using one single model. Conditioning on the target domain label, an auto-encoder-like network with multiple attention connections is trained to translate the input image into the target domain. The attention connections better shuttle the low-level information in the encoder to the decoder, which helps to preserve the structure. A multi-level attention mechanism is also designed in the proposed model to further improve the performance of our model. More specially, the feature maps in the encoder are first squeezed by average pooling and used to output a channel-wise attention mask. The attention mask softly determines which channels of the feature maps are translated and which channels are kept. By enforcing the model to learn a cyclic domain transformation during training, our model does not require paired training data, which greatly improves the versatility to different kinds of data. We experimentally demonstrated the effectiveness of our proposed model on the facial and clothing image translation tasks. The extensive ablations are also conducted to further validate the contribution of the proposed attention module used in our model.

Purpose: Segmental intensity‐modulated radiation therapy (IMRT) delivers a sequence of segments to obtain a desired intensity distribution. Many leaf sequencing algorithms for segmental IMRT have been developed with the aim of reducing delivered monitor units (MUs) and (or) number of segments and, consequently, to reduce the total treatment delivery time. With the development of real‐time detection technology, it is useful to develop leaf sequencing algorithms that consider the detecting probability of markers implanted into or near the target volume.Methods: In this study, the authors defined the concept of marker visibility to denote the marker's detecting probability and proposed a new leaf sequencing algorithm based on the Kamath algorithm. The new algorithm first uses the Kamath algorithm to generate an initial leaf sequence and then performs a series of column transformations to obtain a new leaf sequence that is optimal in terms of MU efficiency and marker visibility. The authors evaluated the performance of the new algorithm with six artificial fields that had randomly generated intensity matrices and 15 clinical fields that had intensity matrices from the IMRT plans for three prostate cancer patients.Results: Compared to the Kamath algorithm, the new algorithm does not increase the total delivered intensity but increases the marker visibility. For the artificial fields, the marker visibility increased from 66.67% to 91.67% for small radiation fields, from 39.29% to 42.86% for medium size fields, and from 31.48% to 37.04% for large fields. For the clinical fields, the marker visibility increased 9%–20% for four fields, 20%–30% for three fields, 30%–40% for two fields, and more than 40% for one field. However, the marker visibility did not change for 4 out of 15 fields.Conclusions: The authors developed a new leaf sequencing algorithm for optimal MU efficiency and marker visibility and also rigorously proved its optimality.