Visual sentiment analysis intends to understand the sentiment evoked by images, which is an important yet challenging field. Existing methods focus on learning single-scale sentiment features from the whole image. However, it has been proved that image emotions are evoked by the sentiment-specific regions of the image, which are related to high-level to low-level sentiment features. To exploit multi-scale sentiment features from both holistic and localized information, we propose a novel end-to-end multi-task framework for joint sentiment- specific regions detection and sentiment classification. Our method contains three components: Multi-Scale Features Extractor, Sentiment-Specific Regions Detection Branch and Sentiment Classification Branch. In the Multi-Scale Features Extractor, by fusing multi-scale features extracted from convolutional neural networks by means of Feature Pyramid Network (FPN) and Adaptively Spatial Feature Fusion (ASFF), the proposed approach first generates a holistic feature carrying more sentiment-related information. Then, the semantic map is automatically discovered from the enhanced holistic feature by adopting attention mechanism in the Sentiment- Specific Regions Detection Branch, which does not require any manual annotations. Finally, the localized and holistic information are adaptively integrated for final sentiment classification in the Sentiment Classification Branch. Extensive experiments on public benchmark datasets demonstrate the robustness and effectiveness of our method in visual sentiment analysis.
Convolutional neural networks have recently demonstrated high-quality reconstruction of single image dehazing. However, existing methods seldom consider the relationship of haze concentration and image depth. In this paper, we propose an end-to-end single image dehazing network called Progressive Guidance Dehazing Network (PGDN), which gradually recovers the clear image from the shallow to deep areas of its hazy image. Our network consists of progressive dehazing blocks, each of which is followed by a guided filtering layer to reinforce the result. Additionally, deep supervisions are added before and after each guided filtering layer, and the supervisions before the guided filtering layers guarantee that the dehazing blocks further incorporate the mutual content information. Experiments on both synthetic and real-world dataset show that our network achieves superior performance over existing methods in quantitative and qualitative evaluations.
Visual object tracking is one of the popular research topics in computer vision. It has a wide range of application scenarios. Although recent approaches based on siamese network have achieved good performance, similar interference and non-real-time speed are still very challenging problems. In this paper, an online Patch Filter Network (OPFNet) is proposed, the online patch filters learned from the target can introduce the local detailed features and avoid the interference of similar objects. In addition, in order to enhance the generalization ability of the tracker trained with small scale dataset, an image mix-up method for augmentation is proposed during offline training process. These improvements are proved to be effective by experiments and can be applied to existed siamese tracking methods
In recent years, quantizing the weights of a deep neural network draws increasing attention in the area of network compression. An efficient and popular way to quantize the weight parameters is to replace a filter with the product of binary values and a real-valued scaling factor. However, the quantization error of such binarization method raises as the number of a filter's parameter increases. To reduce quantization error in existing network binarization methods, we propose group binary weight networks (GBWN), which divides the channels of each filter into groups and every channel in the same group shares the same scaling factor. We binarize the popular network architectures VGG, ResNet and DesneNet, and verify the performance on CIFAR10, CIFAR100, Fashion-MNIST, SVHN and ImageNet datasets. Experiment results show that GBWN achieves considerable accuracy increment compared to recent network binarization methods, including BinaryConnect, Binary Weight Networks and Stochastic Quantization Binary Weight Networks.