Observing the structure of retinal blood vessels can help doctors diagnose the disease of patients, so the accurate segmentation of retinal blood vessels has important research significance. However, there are many problems in retinal vessel segmentation, such as complex and small vascular structure and low image contrast, which lead to low segmentation accuracy. To solve the above problems, this paper proposes an attention-directed adversarial network for retinal vascular segmentation. The purpose is to guide the network to learn useful information for vascular segmentation and ignore useless redundant information. The attention directed generative adversarial networks consist of generator and discriminator. The generator uses U-Net architecture and combines the high-low feature attention modules. The high-low level feature attention modules act on the high-level and low-level feature maps so that the model can strengthen the high-level and low-level spatial features respectively, eliminate redundant information, and guide the model to pay more attention to the vascular foreground information. The batch normalization layer in the generator is also removed to avoid the impact of unstable batch statistics on the segmentation results when the generator is trained in small batches. The discriminator consists of a stack of residual modules, which together with the generator form a conditional generation adversarial network. The experimental results show that the Se, Sp, Acc and AUC of this paper’s method tested on the DRIVE dataset are 82.88%, 97.45%, 95.59%, and 97.86%, respectively, and all the indexes are better than the current mainstream retinal vessel segmentation algorithms.
Person re-identification is a cross camera pedestrian retrieval problem. The data retrieved by pedestrians can be images, videos, and text. The current person re-identification methods are insufficient in expression of pedestrian features and poor robustness, resulting in low model accuracy. This paper proposes a Multi-scale Residual Pooling for person re-identification. ResNet50 is used as the basic network to obtain the multi-scale features. Global average pooling and maximum average pooling are performed on the input features at different network levels. Each group of average pooled and maximum pooled features is subtracted to remove the influence of image background clutter. The subtracted difference features are added to the maximum pooled features to obtain a more discriminative residual pooled fusion feature, making the network focus on the whole body contour of pedestrians and the difference between pedestrians and background. On this basis, triplet loss and cross-entropy loss are combined to optimize the model, and reordering technology is used to optimize the network. The experimental results showed that the Rank1 index of this paper’s method tested on the Market1501 and Duke MTMC-reID datasets reaches 96.41% and 91.43%, respectively, and mAP (mean Average Precision) reaches 94.52% and 89.30%, respectively. which is better than the current mainstream algorithms.
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.