HSVLT: Hierarchical Scale-Aware Vision-Language Transformer for Multi-Label Image Classification

Multi-label image classification is the task of predicting a set of labels corresponding to objects, attributes or other entities present in an image. In this work we propose the Classification Transformer (C-Tran), a general framework for multi-label image classification that leverages Transformers to exploit the complex dependencies among visual features and labels. Our approach consists of a Transformer encoder trained to predict a set of target labels given an input set of masked labels, and visual features from a convolutional neural network. A key ingredient of our method is a label mask training objective that uses a ternary encoding scheme to represent the state of the labels as positive, negative, or unknown during training. Our model shows state-of-the-art performance on challenging datasets such as COCO and Visual Genome. Moreover, because our model explicitly represents the label state during training, it is more general by allowing us to produce improved results for images with partial or extra label annotations during inference. We demonstrate this additional capability in the COCO, Visual Genome, News-500, and CUB image datasets.

Hierarchical GAN-Tree and Bi-Directional Capsules for multi-label image classification

Human visual perception shows good consistency for many multi-label image classification tasks under certain spatial transforms, such as scaling, rotation, flipping and translation. This has motivated the data augmentation strategy widely used in CNN classifier training -- transformed images are included for training by assuming the same class labels as their original images. In this paper, we further propose the assumption of perceptual consistency of visual attention regions for classification under such transforms, i.e., the attention region for a classification follows the same transform if the input image is spatially transformed. While the attention regions of CNN classifiers can be derived as an attention heatmap in middle layers of the network, we find that their consistency under many transforms are not preserved. To address this problem, we propose a two-branch network with an original image and its transformed image as inputs and introduce a new attention consistency loss that measures the attention heatmap consistency between two branches. This new loss is then combined with multi-label image classification loss for network training. Experiments on three datasets verify the superiority of the proposed network by achieving new state-of-the-art classification performance.

Fake news that combines text and images has a better story-telling ability than text-only fake news, making it more deceptive and easier to spread maliciously. Therefore, multi-modal fake news detection has become a new hot topic. There are two main challenges in this task. First, the traditional pre-training model BERT has more parameters and a relatively slow training speed, which limits the level of extracted text features. Second, in the multi-modal form, the fusion process is only a simple splicing of visual and textual features of news, and the obtained multi-modal features are insufficient to express the complementarity between multi-modal data and may have redundant information, potentially leading to biased detection results.In order to solve the above issues, we propose ALB-MCF, using ALBERT and Multi-modal Circulant Fusion(MCF) for fake news detection. ALB-MCF consists of four main modules: a multi-modal feature extractor, a multi-modal feature fusion, a fake news detection and a domain classifier. Specifically, the multi-modal feature extractor innovatively uses a pre-trained ALBERT model to extract text features and a pre-trained VGG19 model to extract visual features. Then, the text features and the visual features are fused into a multi-modal feature representation by MCF, which improves the fusion capability while avoiding an increase in parameters and computational cost. Finally, the multi-modal features are fed into the detector to detect fake news. The role of the domain classifier is mainly to remove event-specific features while retaining shared features between events, thus providing effective detection of emerging events. We have conducted extensive experiments on two real-world datasets. The results demonstrated that our model can handle multi-modal data more effectively, thus improving the accuracy of fake news detection.

In this paper, we present a novel multi-sensor fusion framework for tightly coupled monocular visual-LiDAR odometry and mapping. Compared to previous visual-LiDAR fusion frameworks, our proposed framework leverages more constraints among LiDAR features and visual features and integrates that in a tightly coupled approach. Specifically, the framework starts with a preprocess module which contains LiDAR feature extraction, visual feature extraction and tracking, and visual feature depth recover. Then a frame-to-frame odometry module is established by fusing visual feature tracking and LiDAR feature match between frames, aiming to provide a coarse pose estimation for next module. Finally, to refine the pose and build a multi-modal map, we introduce a multi-modal mapping module that tightly couple multi-modal feature constraints by matching or registering multi-modal features to multi-modal map. In addition, the proposed fusion framework also functions well in sensor-degraded environment (texture-less or structure-less), which increases its robustness. The effectiveness and performance of the proposed fusion framework are demonstrated and evaluated on the public KITTI odometry benchmark, and results show that our proposed fusion framework achieves comparable performance compared with the state-of-the-art visual-LiDAR odometry frameworks.

High dimensional visual features and multimodal based search systems are popular due to their visual retrieval accuracy. But high dimensionality and multimodality negatively affect efficiency and computational complexity. This paper brings the performance advantages of low dimensional visual feature called Compacted Dither Pattern Code (CDPC). CDPC is accompanied with texel properties, spatial colour arrangements, high compactness, and robust feature representation. CDPC feature uses a low computational cost in visual feature extraction which is illustrated using four abstract stages. This feature has proven its performances in our experimental study. In our experiment, CDPC was used alone as a syntactic visual feature for semantic visual depiction and retrieval. Ten irregular shapes based visual concepts were used in visual description. The average precision results generated from ranked retrieval set were used as a metric in evaluation. The retrieval result based on Trecvid 2005 dataset is presented compared to well known low dimensional and high dimensional visual features. Results show the balanced made by CDPC with Bhattacharyya classification in between retrieval accuracy and visual feature space reduction. Competitive advantages of this method over state of art systems are also technically described.

Most existing work on multi-label learning focused on supervised learning which requires manual annotation samples that is labor-intensive, time-consuming and costly. To address such a problem, we present a novel method that incorporates active learning into the semi-supervised learning for multi-label image classification. What’s more, aiming at the curse of dimensionality existing in high-dimensional data, we explore a dimensionality reduction technique with non-negative sparseness constraint to extract a group of features that can completely describe the data and hence make the learning model more efficiently. Experimental results on common data sets validate that the proposed algorithm is relatively effective to improve the performance of the learner in multi-label classification, and the obtained learner is with reliability and robustness after data dimensionality using NNS-DR (Non-Negative Sparseness for Dimensionality Reduction).

Multi-Label Image Classification (MLIC) appro-aches usually exploit label correlations to achieve good performance. However, emphasizing correlation like co-occurrence may overlook discriminative features and lead to model overfitting. In this study, we propose a generic framework named Parallel Self-Distillation (PSD) for boosting MLIC models. PSD decomposes the original MLIC task into several simpler MLIC sub-tasks via two elaborated complementary task decomposition strategies named Co-occurrence Graph Partition (CGP) and Dis-occurrence Graph Partition (DGP). Then, the MLIC models of fewer categories are trained with these sub-tasks in parallel for respectively learning the joint patterns and the category-specific patterns of labels. Finally, knowledge distillation is leveraged to learn a compact global ensemble of full categories with these learned patterns for reconciling the label correlation exploitation and model overfitting. Extensive results on MS-COCO and NUS-WIDE datasets demonstrate that our framework can be easily plugged into many MLIC approaches and improve performances of recent state-of-the-art approaches. The source code is released at https://github.com/Robbie-Xu/CPSD.

Multi-label image classification is a very challenging task, where data are often associated with multiple labels and represented with multiple views. In this paper, we propose a novel multi-view distance metric learning approach to dealing with the multi-label image classification problem. In particular, we attempt to concatenate multiple types of features after learning one optimal distance metric for each view, so as to obtain a better joint representation across multi-view spaces. To preserve the intrinsic geometric structure of the data in the low-dimensional feature space, we introduce a manifold regularization with the adjacency graph being constructed based on all labels. Moreover, we learn another distance metric to capture the dependency of labels, which can further improve the classification performance. Experimental results on publicly available image datasets demonstrate that our method achieves superior performance, compared with the state-of-the-arts.

Learning to rank has recently emerged as an attractive technique to train deep convolutional neural networks for various computer vision tasks. Pairwise ranking, in particular, has been successful in multi-label image classification, achieving state-of-the-art results on various benchmarks. However, most existing approaches use the hinge loss to train their models, which is non-smooth and thus is difficult to optimize especially with deep networks. Furthermore, they employ simple heuristics, such as top-k or thresholding, to determine which labels to include in the output from a ranked list of labels, which limits their use in the real-world setting. In this work, we propose two techniques to improve pairwise ranking based multi-label image classification: (1) we propose a novel loss function for pairwise ranking, which is smooth everywhere and thus is easier to optimize; and (2) we incorporate a label decision module into the model, estimating the optimal confidence thresholds for each visual concept. We provide theoretical analyses of our loss function in the Bayes consistency and risk minimization framework, and show its benefit over existing pairwise ranking formulations. We demonstrate the effectiveness of our approach on three large-scale datasets, VOC2007, NUS-WIDE and MS-COCO, achieving the best reported results in the literature.

In our daily life, the scenes around us are always with multiple labels especially in a smart city, i.e., recognizing the information of city operation to response and control. Great efforts have been made by using Deep Neural Networks to recognize multi-label images. Since multi-label image classification is very complicated, people seek to use the attention mechanism to guide the classification process. However, conventional attention-based methods always analyzed images directly and aggressively. It is difficult for them to well understand complicated scenes. In this paper, we propose a global/local attention method that can recognize an image from coarse to fine by mimicking how human-beings observe images. Specifically, our global/local attention method first concentrates on the whole image, and then focuses on local specific objects in the image. We also propose a joint max-margin objective function, which enforces that the minimum score of positive labels should be larger than the maximum score of negative labels horizontally and vertically. This function can further improve our multi-label image classification method. We evaluate the effectiveness of our method on two popular multilabel image datasets (i.e., Pascal VOC and MS-COCO). Our experimental results show that our method outperforms state-of-the-art methods.

Multi-label image classification (MLIC) deals with assigning multiple labels to each image, a easy task for human being while still a open problem in machine learning. The greatest challenge in MLIC lies in that different target objects in one image keep distinct viewpoints and scales. One effective way is to borrow the label-related information to guide the selection of interesting region, which will act an important role in classification. By leveraging the attention mechanism in transformer, we propose a region-awared transformer to focus on top related regions and neglect background interference. Furthermore, our approach can cope with the positive-negative imbalance by assigning them different exponential decay factors of positive and negative samples separately. Experiments on MS-COCO show a competitive performance against other state-of-the-art methods.

目的 基于计算机的胸腔X线影像疾病检测和分类目前存在误诊率高，准确率低的问题。本文在视觉Transformer（vision Transformer，ViT）预训练模型的基础上，通过迁移学习方法，实现胸腔X线影像辅助诊断，提高诊断准确率和效率。方法 选用带有卷积神经网络（convolutional neural network，CNN）的ViT模型，其在超大规模自然图像数据集中进行了预训练；通过微调模型结构，使用预训练的ViT模型参数初始化主干网络，并迁移至胸腔X线影像数据集中再次训练，实现疾病多标签分类。结果 在IU X-Ray数据集中对ViT迁移学习前、后模型平均AUC（area under ROC curve）得分进行对比分析实验。结果表明，预训练ViT模型平均AUC得分为0.774，与不使用迁移学习相比提升了0.208。并针对模型结构和数据预处理进行了消融实验，对ViT中的注意力机制进行可视化，进一步验证了模型有效性。最后使用Chest X-Ray14和CheXpert数据集训练微调后的ViT模型，平均AUC得分为0.839和0.806，与对比方法相比分别有0.014～0.031的提升。结论 与其他方法相比，ViT模型胸腔X线影像的多标签分类精确度更高，且迁移学习可以在降低训练成本的同时提升ViT模型的分类性能和泛化性。消融实验与模型可视化表明，包含CNN结构的ViT模型能重点关注有意义的区域，高效获取胸腔X线影像的视觉特征。;Objective The chest X-ray-relevant screening and diagnostic method is essential for radiology nowadays. Most of chest X-ray images interpretation is still restricted by clinical experience and challenged for misdiagnose and missed diagnoses. To detect and identify one or more potential diseases in images automatically，it is beneficial for improving diagnostic efficiency and accuracy using computer-based technique. Compared to natural images，multiple lesions are challenged to be detected and distinguished accurately in a single image because abnormal areas have a small proportion and complex representations in chest X-ray images. Current convolutional neural network（CNN）based deep learning models have been widely used in the context of medical imaging. The structure of the CNN convolution kernel has sensitive to local detail information，and it is possible to extract richer image features. However，the convolution kernel cannot be used to get global information，and the features-extracted are restricted of redundant information like its relevance of background， muscles，and bones. The model’s performance in multi-label classification tasks are affected to a certain extent. At present，the vision Transformer（ViT）model has achieved its priorities in computer vision-related tasks. The ViT can be used to capture information simultaneously and effectively for multiple regions of the entire image. However，it is required to use large-scale dataset training to achieve good performance. Due to some factors like patient privacy and manual annotate costs，the size of the chest X-ray image data set has been limited. To reduce the model's dependence on data scale and improve the performance of multi-label classification，we develop the CNN-based ViT pre-training model in terms of the transfer learning method for diagnosis-assisted of chest X-ray image and multi-label classification. Method The CNN-based ViT model is pre-trained on a huge scale ground truth dataset，and it is used to obtain the initial parameters of the model. The model structure is fine-tuned according to the features of chest X-ray dataset. A 1×1 convolution layer is used to convert the chest X-ray images channels between 1 to 3. The number of output nodes of the linear layer in the classifier is balanced from 1 000 to the number of chest X-ray classification labels，and the Sigmoid is used as an activation function. The parameters of the backbone network are initialized in terms of the pre-trained ViT model parameters，and it is trained in the chest X-ray dataset after that to complete multi-label classification. The experiment is configured of Python3. 7 and PyTorch1. 8 to construct the model and RTX3090 GPU for training. Stochastic gradient descent（SGD）optimizer，binary cross-entropy（BCE）loss function，an initial learning rate of 1E-3，the cosine annealing learning rate decay are used. For training，each image is scaled to a size of 512×512 pixels，and a 224×224 pixels area and it is then cropped in random as the model input，and data augmentation is performed randomly by some of the flipping，perspective transformation， shearing，translation，zooming，and changing brightness. For testing，the chest X-ray image is scaled to 256×256 pixels and center crop a 224×224 area to input the trained model. Result The experiment is performed on the IU X-Ray，which is a small-scale chest X-ray dataset. This model is evaluated in quantitative using the average of area under ROC curve （AUC）scores across all classification labels. The results show that the average AUC score of the pre-trained ViT model is 0. 774. The accuracy and training efficiency of the non-pre-trained ViT model is dropped significantly. The average AUC score is reached to 0. 566 only，which is 0. 208 lower. In addition，the attention mechanism heat map is generated based on the ViT model，which can strengthen the interpretability of the model. A series of ablation experiments are carried out for data augmentation，model structure，and batch size design. The fine-tuned ViT model is trained on the Chest-Ray14 and CheXpert dataset as well. The average AUC score is reached to 0. 839 and 0. 806，which is optimized by 0. 014 and 0. 031. Conclusion A pre-trained ViT model is used for the multi-label classification of chest X-ray images via transfer learning. The experimental results illustrate that the ViT has its stronger multi-label classification performance in chest Xray images，and its attention mechanism is beneficial for lesions precision-focused like the interior of the chest cavity and the heart. Transfer learning is potential to improve the classification performance and model generalization of the ViT in small-scale datasets，and the training cost is reduced greatly. Ablation experiments demonstrate that the incorporated model of CNN and Transformer has its priority beyond single-structure model. Data enhancement and the batch size cutting can improve the performance of the model，but smaller scale of batch is still interlinked to longer training span. To improve the model's ability，we predict that future research direction can be focused on the extraction for complex disease and highlevel semantic information，such as their small lesions，disease location，and severity.

Much progress has been achieved during the past two decades in audio-visual automatic speech recognition (AVASR). However, challenges persist that hinder AVASR deployment in practical situations, most notably, robust and fast extraction of visual speech features. We review our efforts in overcoming this problem, based on an appearance-based visual feature representation of the speaker's mouth region. We cover three topics in particular. Firstly, we discuss AVASR in realistic, visually challenging domains, where lighting, background, and head-pose vary significantly. To enhance visual-front-end robustness in such environments, we employ an improved statistical-based face detection algorithm that significantly outperforms our baseline scheme. However, visual-only recognition remains inferior to visually clean (studio-like) data, thus demonstrating the importance of accurate mouth region extraction. We then consider a wearable audio-visual sensor to capture the mouth region directly, thus eliminating face detection. Its use improves visual-only recognition, even over full-face videos recorded in the studio-like environment. Finally, we address the speed issue in visual feature extraction, by discussing our real-time AVASR prototype implementation. The reported progress demonstrates the feasibility of practical AVASR.

<p>Visual grounding locates target objects or areas in the image based on natural language expression. Most current methods extract visual features and text embeddings independently, and then carry out complex fusion reasoning to locate target objects mentioned in the query text. However, such independently extracted visual features often contain many features that are irrelevant to the query text or misleading, thus affecting the subsequent multimodal fusion module, and deteriorating target localization. This study introduces a combined network model based on the transformer architecture, which realizes more accurate visual grounding by using query text to guide visual feature generation and multi-stage fusion reasoning. Specifically, the visual feature generation module reduces the interferences of irrelevant features and generates visual features related to query text through the guidance of query text features. The multi-stage fused reasoning module uses the relevant visual features obtained by the visual feature generation module and the query text embeddings for multi-stage interactive reasoning, further infers the correlation between the target image and the query text, so as to achieve the accurate localization of the object described by the query text. The effectiveness of the proposed model is experimentally verified on five public datasets and the model outperforms state-of-the-art methods. It achieves an improvement of 1.04%, 2.23%, 1.00% and +2.51% over the previous state-of-the-art methods in terms of the top-1 accuracy on TestA and TestB of the RefCOCO and RefCOCO+ datasets, respectively.</p> <p> </p>