Accurately identifying drug-target interactions (DTIs) is crucial for accelerating drug discovery. While computational methods offer efficient alternatives to experimental validation, most existing models struggle to integrate diverse structural and sequential characteristics of drugs and proteins within a unified framework. We propose UMSDTI, a deep learning model that captures molecular structural information from sequence, atom, and bond perspectives, and encodes proteins via integrating sequence and pre-trained modelderived contact maps. A co-attention fusion module jointly learns drug-target dependencies for enhanced prediction. Experiments on five benchmark datasets demonstrate that UMSDTI outperforms state-of-the-art methods in both prediction performance and biological insight. Source codes are available at https://anonymous.4open.science/r/UMSDTI.

Prediction of drug-target interactions (DTI) plays a vital role in drug discovery and repositioning. Although extensive efforts have been invested in improving DTI prediction performance, existing methods still struggle with the cold-start problem where there are few or no interactions between new drugs and targets. In this study, we propose an Information Entropy Guided Multimodal Contrastive Learning framework (IEGMCL) to address the cold-start problem. Specifically, we introduce an information entropy guided gated fusion mechanism which dynamically adjusts the weights of drug graph structures and sequence features, improving robustness to modality imbalance under cold-start scenarios. Furthermore, we introduce a contrastive learning method with hard negative sample mining to mitigate inter-modality discrepancies and semantic drift, boosting the generalization under cold-start scenario. Experimental results on the benchmark datasets demonstrate that IEGMCL achieves superior prediction performance to existing mainstream methods under three cold-start settings: unseen drugs, unseen targets, both unseen drugs and unseen targets.

Visual Text Question Answering (VTQA) is a challenging task that requires answering questions pertaining to visual content by combining image understanding and language comprehension. The main objective is to develop models that can accurately provide relevant answers based on complementary information from both images and text, as well as the semantic meaning of the question. Despite ongoing efforts, the VTQA task presents several challenges, including multimedia alignment, multi-step cross-media reasoning, and handling open-ended questions. This paper introduces a novel generative framework called VTQAGen, which leverages a Multi- modal Attention Layer to combine image-text pairs and question inputs, as well as a BART-based model for reasoning and entity extraction from both images and text. The framework incorporates a step-based ensemble method to enhance model performance and generalization ability. VTQAGen utilizes an encoder-decoder generative model based on BART. Faster R-CNN is employed to extract visual regions of interest, while BART's encoder is modified to handle multi-modal interaction. The decoder stage utilizes the shift-predict approach and introduces step-based logits fusion to improve stability and accuracy. In the experiments, the proposed VTQAGen demonstrates superior performance on the testing set, securing second place in the ACM Multimedia Visual Text Question Answer Challenge.

The discovery of drug-target interactions (DTIs) is a pivotal process in pharmaceutical development. Computational approaches are a promising and efficient alternative to tedious and costly wet-lab experiments for predicting novel DTIs from numerous candidates. Recently, with the availability of abundant heterogeneous biological information from diverse data sources, computational methods have been able to leverage multiple drug and target similarities to boost the performance of DTI prediction. Similarity integration is an effective and flexible strategy to extract crucial information across complementary similarity views, providing a compressed input for any similarity-based DTI prediction model. However, existing similarity integration methods filter and fuse similarities from a global perspective, neglecting the utility of similarity views for each drug and target. In this study, we propose a Fine-Grained Selective similarity integration approach, called FGS, which employs a local interaction consistency-based weight matrix to capture and exploit the importance of similarities at a finer granularity in both similarity selection and combination steps. We evaluate FGS on five DTI prediction datasets under various prediction settings. Experimental results show that our method not only outperforms similarity integration competitors with comparable computational costs, but also achieves better prediction performance than state-of-the-art DTI prediction approaches by collaborating with conventional base models. Furthermore, case studies on the analysis of similarity weights and on the verification of novel predictions confirm the practical ability of FGS.

In recent years, Convolutional Neural Networks (CNNs) have achieved excellent results in the study of single image super-resolution. However, super-resolution algorithms based on CNNs still face serious challenges, such as poor detail reconstruction, numerous parameters, and difficulty of training. A Residual Dense Information Distillation Network (RD-IDN) is proposed in this paper which uses dense skip connections and residual structure to solve the problems of difficult training and low utilization of features in Information Distillation Network. Experimental results show that the proposed method is superior to many other Super Resolution algorithms in terms of reconstruction performance and computational comsumption.

Lane detection task is the most necessary part in modern intelligent driving technology. However, there still have many challenge need to be conquered. In this paper, we proposed a three-branch neural network, in this framework, there are several new techniques are used. Including multi-branch, lightweight module, feature recalibration and decoder module, named Three Branch Net. Furthermore, A new dataset have been used, which includes much more complex situation and more close to the real world. Compared with other newest method, experiment results shows the proposed approach is the most effective method in complex road conditions task.

The recent years witnessed great improvements in systems of region-based face detection. However, the variations in occlusion, scale, illumination, pose and facial expressions make face detection in the wild still a challenge to be solved. In this paper, a Region-based Fully Convolutional Networks (R-FCN) based deep face detection framework is proposed. Several new techniques are utilized in our framework, including Deformable Convolutional Networks (DCN), Feature Pyramid Networks (FPN) and Focal Loss. Experiment results on three common challenging face detection benchmarks, FDDB, AFW and WIDER FACE, show the proposed approach is robust and performance outperforms most of previous methods, especially for addressing heavy occlusion, part deformation and complex perspective.

Nowadays, more and more methods have been proposed to solve the problem of face detection based on computer implementation. Due to the variations in background, illumination, pose and facial expressions, the problem of machine face detection is complex. Recently, deep learning approaches achieve an impressive performance on face detection. In this paper, a model named Multi-Scale Fusion Convolutional Neural Network (MSF-CNN) is proposed to train the face detector. The model is trained by Convolutional Neural Network and detecting is based on the Viola & Jones detector's sliding windows structure. Particularly, in the process of feature extraction, we adopt the design of multi-scale feature fusion with different scale convolution kernels. The results are as follows: First, the fusion of multi-scale features are rich in the characteristics of learning, and the classification accuracy is higher than the single-scale. Second, we decrease the model of complexity compared with existed methods of the cascaded CNN. Third, we achieve end-to-end learning compared with cascaded separate training. Meanwhile, the proposed model has showed that the performance of results outperforms the previous methods in some well-known face detection benchmark datasets.

Traditional digital display devices, due to their hardware limitations, cannot represent the whole range of luminance in High Dynamic Range (HDR) images. In order to solve this incompatible problem, many tone mapping techniques were introduced to reproduce HDR images presently. Unlike one of the traditional work of art in [13], this paper proposes a fast and multi-scale decomposition based tone mapping algorithm using the Improved Local Extrema (ILE) filter depended on the correlation of pixels. The reason of using ILE filter is due to the fact that it is able to decrease the time-consuming without noticeable image quality deterioration. Firstly, the ILE filters of variant scales are utilized to dispose the input HDR image into a series of base images under different scales. Secondly, multi-scale decomposition is utilized to obtain detail images with variable scales from the aforementioned base images. Finally, both of the base and detail images are converted into an initial compressed image to generate a Low Dynamic Range (LDR) image. Experimental results show that the proposed algorithm outperforms previous methods for reconstructing the real scene of HDR images, especially for its fast running time compared with the traditional work using Local Extrema (LE) filter.

Multilabel classification is prevalent in many real-world applications where data instances may be associated with multiple labels simultaneously. In multilabel classification, exploiting label correlations is an essential but nontrivial task. Most of the existing multilabel learning algorithms are either ineffective or computationally demanding and less scalable in exploiting label correlations. In this paper, we propose a co-evolutionary multilabel hypernetwork (Co-MLHN) as an attempt to exploit label correlations in an effective and efficient way. To this end, we firstly convert the traditional hypernetwork into a multilabel hypernetwork (MLHN) where label correlations are explicitly represented. We then propose a co-evolutionary learning algorithm to learn an integrated classification model for all labels. The proposed Co-MLHN exploits arbitrary order label correlations and has linear computational complexity with respect to the number of labels. Empirical studies on a broad range of multilabel data sets demonstrate that Co-MLHN achieves competitive results against state-of-the-art multilabel learning algorithms, in terms of both classification performance and scalability with respect to the number of labels.

A robotic fish should be designed to swim with the obstacle avoidance capability in the real world. In this paper, a neuro-fuzzy control method is proposed for a multi-joint robotic fish using 3D printing technology. For this method, we can use the general infrared sensors to measure the existence of obstacles and the distance from the robotic fish to the obstacles. With the inference and learning ability of the neuro-fuzzy control system, the robotic fish can move away from obstacles freely. Moreover, in order to solve the waterproof problem and facilitate the manufacturing, the 3D printing technology is introduced to construct a real 3D printed robotic fish. The simulations and real experiments demonstrate the merits and practical applicability of the proposed method.

In traditional image retrieval techniques, the query results are severely affected when the images of varying illumination and scale, as well as occlusion and corrosion. Seeking to solve this problem, this paper proposed a novel multi-feature fusion and sparse coding based framework for image retrieval. In the framework, firstly, inherent features of an image are extracted, and then dictionary learning method is utilized to construct them to be dictionary features. Finally, the proposed framework introduces sparse representation model to measure the similarity between two images. The merit is that a feature descriptor is coded as a sparse linear combination with respect to dictionary feature so as to achieve efficient feature representation and robust similarity measure. In order to check the validity of the framework, this paper conducted two groups of experiments on Corel-1000 image dataset and the Stirmark benchmark based database respectively. Experimental results show that the proposed framework is much more effective than the state-of-the-art methods not only in traditional image dataset but also in varying image dataset.

Conventional digital display devices, due to their hardware limitations, can't represent the whole range of luminance in High Dynamic Range (HDR) images. In order to solve this incompatible problem, many tone mapping techniques were introduced to reproduce HDR images. Unlike the traditional methods applied in standard scale space, this paper proposes a novel affine invariant features-based tone mapping algorithm in affine Gaussian scale space. The reason of using this scale space is due to the fact that it is able to extract anisotropic feature regions in addition to traditional isotropic feature regions. Firstly, the proposed method extracts the anisotropic features from HDR images and reforms them to be isotropic by Fitting & Affine transformation. Then, dodging-and-burning processing is utilized to obtain base layer of HDR images. Finally, two-scale edge-preserving decomposition is employed to generate detail layer of HDR images and combine two layers to produce output images. Experimental results show that the proposed algorithm outperforms previous methods for reproducing the real scene of HDR images, especially for large perspective and scale transformed images which contain considerable anisotropic feature regions.

Hyper networks consist of a large number of hyper edges that represent high-order features sampled from training sets. The order of hyper edges is an important parameter of a hyper network model and influences the performance of the hyper network classification system. Previous studies determine the parameter by the artificial exhaustive search method before evolutionary learning. Not only is the approach time-consuming, but also the traditional hyper network lacks generalization. In this study, a genetic algorithm is employed to optimize the order of hyper networks. The proposed method is tested on the acute leukemia and the colon cancer dataset. Experimental results show that the proposed approach can find the global optimal order automatically. Also, a comparative study on five classification algorithms shows that the improved hyper network model achieves a comparable classification performance.

Hyper network is a probabilistic graphic model of learning and memory inspired by biomolecular networks, which is very useful for discovering higher-order correlations among multiple attributes. However, as many traditional machine learning algorithms, hyper networks may bias towards the majority class, thus producing poor predictive accuracy over the minority class when learining with imbalacned datasets. In this paper, three hyper network-based models, namely ensemble of cost-sensitive hyper networks (EN-CS-HN), ensemble of cost-sensitive hyper networks with under-sampling (EN-CS-HN-UNDE), and ensemble of cost-sensitive hyper networks with synthetic minority over-sampling technique (EN-CS-HN-SMOTE) are proposed respectively. To examine the performance of the proposed schemes, we conduct experiments on ten imbalanced datasets collected from UCI machine learning repository, wherein the proposed methods are compared with various state-of-the-art approaches using three metrics: G-Mean, F-Measure and area under the receiver operating characteristic curve (AUC-ROC). Experimental results show that the proposed methods are able to surpass or match the previously known best algorithms on most of the ten datasets.