CE-DCVSI: Multimodal relational extraction based on collaborative enhancement of dual-channel visual semantic information

Visual information is taking a predominant place in our society. With the advent of digital technologies, visual information will find new applications in domains ranging from communication, commerce to entertainment. New functionalities will be required that permit an extended interaction with the visual information. In particular, this is the case for digital television and the related problem of video sequence compression. The extent of possible interactions depends on the manner in which the visual information is represented. Up to now, the canonical representation is used. Also referred to as the waveform representation, it is a technical artifact of the image capture procedure. This representation severely restricts the functionalities available to the end-user. The latter is unable to freely manipulate or customize the received visual information. In this dissertation, we propose to describe the visual information through a semantically meaningful representation. This representation directly derives from the scene content, which is decomposed in terms of its constituting objects. Consequently, the viewing process is totally disconnected from the image capture procedure. This permits full interactivity with the visual information, leading to enhanced functionalities for the end-user. The essence of this dissertation is to automatically define the objects forming the scene arid to automatically track them through the video sequence. Two main issues are identified: the initial segmentation of the objects and their tracking. The first issue deals with segmenting the scene into its constituent objects. This has to be performed on the basis of the information available in two consecutive frames. In order to solve the problem, a split-and-merge approach is used. First, the image is segmented into small, spatio-temporally homogeneous regions. This is achieved through a top-down approach where spatial, temporal and change information is combined. These regions are used as a starting point to define the objects forming the scene. A bottom-up approach is used which iteratively merges the regions. The propensity of the regions to form an object is assessed in terms of both spatial and temporal information. The second issue deals with segmenting and tracking the objects in the successive frames. The coherence between the successive segmentations is ensured by using past and current information. Also, the different objects composing the scene are identified throughout the video sequence. This identification relies on temporal, spatial and spatio-temporal features of the objects. The proposed representation of the visual information finds a natural application in video coding. A second generation video coding scheme is presented which combines compression efficiency with extended functionalities. In particular, scalable coding is achieved in terms of both scene content and object coding quality.

Deep learning is skilled at learning representation from raw data, which are embedded in the semantic space. Traditional multimodal networks take advantage of this, and maximize the joint distribution over the representations of different modalities. However, the similarity among the representations are not emphasized, which is an important property for multimodal data. In this paper, we will introduce a novel learning method for multimodal networks, named as Semantic Similarity Learning (SSL), which aims at training the model via enhancing the similarity between the high-level features of different modalities. Sets of experiments are conducted for evaluating the method on different multimodal networks and multiple tasks. The experimental results demonstrate the effectiveness of SSL in keeping the shared information and improving the discrimination. Particularly, SSL shows its ability in encouraging each modality to learn transferred knowledge from the other one when faced with missing data.

Multimodal sentiment analysis based on disentangled representation learning and cross-modal-context association mining

During previous couple of years, the World Wide Web (WWW) has become an extremely well-liked information source. To successfully utilize the vast quantity of information that the web provides, we want an effective way to explore it. Image data is much more voluminous than textual data, and visual information cannot be indexed by traditional strategies developed for indexing textual information. Therefore, Content-Based Image Retrieval (CBIR) has received an excellent deal of interest within the research community. A CBIR system operates on the visible features at low-level of a user's input image that makes it troublesome for the users to devise the input and additionally doesn't offer adequate retrieval results. In CBIR system, the study of the useful representation of features and appropriate similarity metrics is extremely necessary for improving the performance of retrieval task. Semantic gap has been the main issue which occurs between image pixels at low-level and semantics at high-level interpreted by humans. Among varied methods, machine learning (ML) has been explored as a feasible way to reduce the semantic gap. Inspired by the current success of deep learning methods for computer vision applications, in this paper, we aim to confront an advance deep learning method, known as Convolutional Neural Network (CNN), for studying feature representations and similarity measures. In this paper, we explored the applications of CNNs towards solving classification and retrieval problems. For retrieval of similar images, we agreed on using transfer learning to apply the GoogleNet deep architecture to our problem. Extracting the last-but-one fully connected layer from the retraining of GoogleNet CNN model served as the feature vectors for each image, computing Euclidean distances between these feature vectors and that of our query image to return the closest matches in the dataset.

Multimodal sentiment analysis aims to utilize multisource information such as text, speech and vision to more comprehensively and accurately identify an individual’s emotional state. However, existing methods still face challenges in practical applications, including modality heterogeneity, insufficient expressive power of non-verbal modalities, and low fusion efficiency. To address these issues, this paper proposes a Modality Enhanced Multimodal Integration Model (MEMMI). First, a modality enhancement module is designed to leverage the semantic guidance capability of the text modality, enhancing the feature representation of non-verbal modalities through a multihead attention mechanism and a dynamic routing strategy. Second, a gated fusion mechanism is introduced to selectively inject speech and visual information into the dominant text modality, enabling robust information completion and noise suppression. Finally, a combined attention fusion module is constructed to synchronously fuse information from all three modalities within a unified architecture, hile a multiscale encoder is used to capture feature representations at different semantic levels. Experimental results on three benchmark datasets—CMU-MOSEI, CMU-MOSI, and CH-SIMS—demonstrate the superiority of the proposed model. On CMU-MOSI, it achieves an Acc-7 of 45.91, with binary accuracy/F1 of 82.86/84.60, MAE of 0.734, and Corr of 0.790, outperforming TFN and MulT by a large margin. On CMU-MOSEI, the model reaches an Acc-7 of 54.17, Acc-2/F1 of 83.69/86.02, MAE of 0.526, and Corr of 0.779, surpassing all baselines, including ALMT. On CH-SIMS, it further achieves 41.88, 66.52, and 77.68 in Acc-5/Acc-3/Acc-2, with F1 of 77.85, MAE of 0.450, and Corr of 0.594, establishing new state-of-the-art performance across datasets. These results confirm that MEMMI achieves state-of-the-art performance across multiple metrics. Furthermore, ablation studies validate the effectiveness of each module in enhancing modality representation and fusion efficiency.

Maintenance manuals are crucial information sources for maintenance and repair. Prior studies explored factual knowledge extraction from textual documents. However, maintenance knowledge in manuals is more task‐centric rather than factual knowledge and often documented in an unstructured Portable Document Format (PDF), posing challenges for knowledge extraction. Addressing this, this research develops effective methods to extract task‐centric maintenance knowledge from unstructured PDF manuals. A new Task‐centric Knowledge Graph (TCKG) schema centralized on maintenance task components (MTCs) is proposed to address the need for structured knowledge representation. A method (Heterogeneous Graph‐based Method, HGM) for knowledge extraction is then proposed, which is enhanced by incorporating visual and spatial information. In the experiments, the proposed HGM exhibits robust performance in the knowledge extraction process, surpassing the baseline Graph‐based Interaction Model with a Tracker (GIT) method in MTCs extraction by 13.3%, and the baseline Translate Embedding (TransE) method in MTCs' relation extraction by 3.8%. A series of ablation studies also prove that including visual and spatial information through the proposed method can improve the relation extraction performance by over 10%. This research supplies valuable insights for future developments in information extraction from maintenance manuals.

Exploiting the relevant speech information that is embedded in facial images has been a significant research topic in recent years, because it has provided complementary information to acoustic signals for a wide range of automatic speech recognition (ASR) tasks. Visual information is particularly important in many real applications where acoustic signals are corrupted by environmental noises. This chapter reviews the most recent advances in feature extraction and representation for Visual Speech Recognition (VSR). In comparison with other surveys published in the past decade, this chapter presents a more up-to-date survey and highlights the strengths of two newly developed approaches (i.e., graph-based learning and deep learning) for VSR. In particular, we summarise the methods of using these two techniques to overcome one of the most challenging difficulties in this area-that is, how to automatically learn good visual feature representations from facial images to replace the widely used handcrafted features. This chapter concludes by discussing potential visual feature representation solutions that may overcome the remaining challenges in this domain.

Deep learning models can be applied successfully in real-work problems; however, training most of these models requires massive data. Recent methods use language and vision, but unfortunately, they rely on datasets that are not usually publicly available. Here we pave the way for further research in the multimodal language-vision domain for radiology. In this paper, we train a representation learning method that uses local and global representations of the language and vision through an attention mechanism and based on the publicly available Indiana University Radiology Report (IU-RR) dataset. Furthermore, we use the learned representations to diagnose five lung pathologies: atelectasis, cardiomegaly, edema, pleural effusion, and consolidation. Finally, we use both supervised and zero-shot classifications to extensively analyze the performance of the representation learning on the IU-RR dataset. Average Area Under the Curve (AUC) is used to evaluate the accuracy of the classifiers for classifying the five lung pathologies. The average AUC for classifying the five lung pathologies on the IU-RR test set ranged from 0.85 to 0.87 using the different training datasets, namely CheXpert and CheXphoto. These results compare favorably to other studies using UI-RR. Extensive experiments confirm consistent results for classifying lung pathologies using the multimodal global local representations of language and vision information.

Recent works of music representation learning mainly focus on learning acoustic music representations with unlabeled audios or further attempt to acquire multi-modal music representations with scarce annotated audio-text pairs. They either ignore the language semantics or rely on labeled audio datasets that are difficult and expensive to create. Moreover, merely modeling semantic space usually fails to achieve satisfactory performance on music recommendation tasks since the user preference space is ignored. In this paper, we propose a novel Hierarchical Two-stage Contrastive Learning (HTCL) method that models similarity from the semantic perspective to the user perspective hierarchically to learn a comprehensive music representation bridging the gap between semantic and user preference spaces. We devise a scalable audio encoder and leverage a pre-trained BERT model as the text encoder to learn audio-text semantics via large-scale contrastive pre-training. Further, we explore a simple yet effective way to exploit interaction data from our online music platform to adapt the semantic space to user preference space via contrastive fine-tuning, which differs from previous works that follow the idea of collaborative filtering. As a result, we obtain a powerful audio encoder that not only distills language semantics from the text encoder but also models similarity in user preference space with the integrity of semantic space preserved. Experimental results on both music semantic and recommendation tasks confirm the effectiveness of our method.

In the context of the rapid development of multimedia and information technology, machine translation plays an indispensable role in cross-border e-commerce between China and Japan. However, due to the complexity and diversity of natural languages, a single neural machine translation model tends to fall into local optimality, leading to poor accuracy. To solve this problem, this paper proposes a general multimodal machine translation model based on visual information. First, visual information and text information are used to generate a visual representation of perceptual text information. Then, the two modal information are encoded separately, and the proportion of visual information in the whole multimodal information is controlled by a gating network. Finally, experiments are conducted on the image description datasets MSCOCO, Flickr30k, and video dataset VATEX, respectively. The results show that the algorithm in this paper achieves the best performance on both the BLEU and METEOR evaluation metrics.

Having access to multi-modal cues (e.g. vision and audio) empowers some cognitive tasks to be done faster compared to learning from a single modality. In this work, we propose to transfer knowledge across heterogeneous modalities, even though these data modalities may not be semantically correlated. Rather than directly aligning the representations of different modalities, we compose audio, image, and video representations across modalities to uncover richer multi-modal knowledge. Our main idea is to learn a compositional embedding that closes the cross-modal semantic gap and captures the task-relevant semantics, which facilitates pulling together representations across modalities by compositional contrastive learning. We establish a new, comprehensive multi-modal distillation benchmark on three video datasets: UCF101, ActivityNet, and VGGSound. Moreover, we demonstrate that our model significantly outperforms a variety of existing knowledge distillation methods in transferring audio-visual knowledge to improve video representation learning. Code is released here: https://github.com/yanbeic/CCL.

Nowadays massive amount of images and texts has been emerging on the Internet, arousing the demand of effective cross-modal retrieval such as text-to-image search and image-to-text search. To eliminate the heterogeneity between the modalities of images and texts, the existing subspace learning methods try to learn a common latent subspace under which cross-modal matching can be performed. However, these methods usually require fully paired samples (images with corresponding texts) and also ignore the class label information along with the paired samples. This may inhibit these methods from learning an effective subspace since the correlations between two modalities are implicitly incorporated. Indeed, the class label information can reduce the semantic gap between different modalities and explicitly guide the subspace learning procedure. In addition, the large quantities of unpaired samples (images or texts) may provide useful side information to enrich the representations from learned subspace. Thus, in this paper we propose a novel model for cross-modal retrieval problem. It consists of 1) a semi-supervised coupled dictionary learning step to generate homogeneously sparse representations for different modalities based on both paired and unpaired samples; 2) a coupled feature mapping step to project the sparse representations of different modalities into a common subspace defined by class label information to perform cross-modal matching. Experiments on a large scale web image dataset MIRFlickr-1M with both fully paired and unpaired settings show the effectiveness of the proposed model on the cross-modal retrieval task.

Attention maintains task-relevant information in working memory (WM) in an active state. We investigated whether the attention-based maintenance of stimulus representations that were encoded through different modalities is flexibly controlled by top-down mechanisms that depend on behavioral goals. Distinct components of the ERP reflect the maintenance of tactile and visual information in WM. We concurrently measured tactile (tCDA) and visual contralateral delay activity (CDA) to track the attentional activation of tactile and visual information during multimodal WM. Participants simultaneously received tactile and visual sample stimuli on the left and right sides and memorized all stimuli on one task-relevant side. After 500 msec, an auditory retrocue indicated whether the sample set's tactile or visual content had to be compared with a subsequent test stimulus set. tCDA and CDA components that emerged simultaneously during the encoding phase were consistently reduced after retrocues that marked the corresponding (tactile or visual) modality as task-irrelevant. The absolute size of cue-dependent modulations was similar for the tCDA/CDA components and did not depend on the number of tactile/visual stimuli that were initially encoded into WM. Our results suggest that modality-specific maintenance processes in sensory brain regions are flexibly modulated by top-down influences that optimize multimodal WM representations for behavioral goals.

Most existing methods on affective video content analyses are dedicated to single media, either visual content or audio content and few attempts for combined analysis of the two media signals are made. In this paper, we employ a cross-modal embedding learning approach to learn the compact feature representations of different modalities that are discriminative for analyzing the emotion attributes of the video. Specifically, we introduce inter-modal similarity constraints and intra-modal similarity constraints to promote the joint embedding learning procedure for obtaining the robust features. In order to capture cues in different grains, global and local features are extracted from both visual and audio signals, thereafter a unified framework consisting with global and local features embedding networks is built for affective video content analyses. Experiments show that our proposed approach significantly outperforms the state-of-the-art methods and demonstrate the effectiveness of our approach.

Visual information retrieval (VIR) is an active and vibrant research area, which attempts at providing means for organizing, indexing, annotating, and retrieving visual information (images and videos) form large, unstructured repositories. The goal of VIR is to retrieve the highest number of relevant matches to a given query (often expressed as an example image and/or a series of keywords). In its early years (1995-2000) the research efforts were dominated by content-based approaches contributed primarily by the image and video processing community. During the past decade, it was widely recognized that the challenges imposed by the semantic gap (the lack of coincidence between an image's visual contents and its semantic interpretation) required a clever use of textual metadata (in addition to information extracted from the image's pixel contents) to make image and video retrieval solutions efficient and effective. The need to bridge (or at least narrow) the semantic gap has been one of the driving forces behind current VIR research. Additionally, other related research problems and market opportunities have started to emerge, offering a broad range of exciting problems for computer scientists and engineers to work on. In this tutorial, we present an overview of visual information retrieval (VIR) concepts, techniques, algorithms, and applications. Several topics are supported by examples written in Java, using Lucene (an open-source Java-based indexing and search implementation) and LIRE (Lucene Image REtrieval), an open-source Java-based library for content-based image retrieval (CBIR) written by Mathias Lux.After motivating the topic, we briefly review the fundamentals of information retrieval, present the most relevant and effective visual descriptors currently used in VIR, the most common indexing approaches for visual descriptors, the most prominent machine learning techniques used in connection with contemporary VIR solutions, as well as the challenges associated with building real-world, large scale VIR solutions, including a brief overview of publicly available datasets used in worldwide challenges, contests, and benchmarks. Throughout the tutorial, we integrate examples using LIRE, whose main features and design principles are also discussed. Finally, we conclude the tutorial with suggestions for deepening the knowledge in the topic, including a brief discussion of the most relevant advances, open challenges, and promising opportunities in VIR and related areas.The tutorial is primarily targeted at experienced Information Retrieval researchers and practitioners interested in extending their knowledge of document-based IR to equivalent concepts, techniques, and challenges in VIR. The acquired knowledge should allow participants to derive insightful conclusions and promising avenues for further investigation.