Search

Homography estimation is a basic and important step in many computer vision tasks. Traditional homography estimation methods are based on feature point matching, which are difficult to work in weak texture images. Deep learning has been applied to homography estimation to improve its robustness, but the existing methods do not consider the multi-scale problem caused by object scale differences, resulting in limited accuracy. To solve the above problems, this paper proposes a multi-scale residual network for homography estimation. The network can extract the multi-scale feature of the image, and used the Multi-Scale Feature Fusion Module to effectively fuse the features. In addition, it further reduced the difficulty of network optimization by estimating the four-corner normalized offset. Experiments on MS-COCO dataset showed that the average corner error of this method was only 0.788 pixels, which achieved sub-pixel accuracy, and can maintain high accuracy in 99% of cases. Due to the comprehensive utilization of multi-scale features and easier to optimize, this method had significantly improved accuracy and stronger robustness. [ABSTRACT FROM AUTHOR]

Intelligent question answering based on reading comprehension refers to letting computers read and comprehend texts like humans, extracts the text information and answers corresponding questions. The pre-training model RoBERTa-wwm-ext uses the extracted original fragments as the answers to the questions, but this method can’t solve the two situations that the answer fragments don’t exist in the original text or need to reply to the original text after summarizing. The pre-training model is used for generative model training, which can solve the problems that need to summarize the original text to a certain extent. Therefore, this paper improved the method of only using RoBERTa-wwm-ext model to extract answers. On this basis, it integrated the generative question answering model based on RAG model to answer questions that could not be handled by Roberta-wwm-ext and other extraction models. At the same time, this paper absorbed the advantages of PGN model, improved RAG model, and obtained RPGN sub model, which could make better use of reading and understanding articles to generate reasonable answers. Therefore, this paper proposed a fusion model of RPR（RAG, PGN, RoBERTa-wwm-ext）,which could be used to deal with both extractive question task and generative question answering task at the same time. [ABSTRACT FROM AUTHOR]

Joint extraction of entities and relations from unstructured text is an important task in information extraction. The existing methods have achieved considerable performance, but are still subject to some inherent limitations, such as error propagation, redundancy of relation prediction, inability to solve the problem of relations overlap, etc. For this reason, this paper proposed a joint entity relationship extraction model BSGB( BiLSTM + SDA-GAT + BiGCN) based on graph neural network. BSGB was a two-stage predicting process. The first stage of this model extended the semantic dependency analysis to the semantic dependency graph, and proposed a graph attention network to integrate the semantic dependency graph(SDA-GAT). By stacking BiLSTM and SDA-GAT, it extracted sentence sequence and local dependent features, and performed entity span detection and preliminary relationship prediction. In the second stage, it constructed the relation-weighted GCN, which further modeled the interaction between entities and relations, and completed the final extraction of entity-relationship triples. The experimental results on the NYT dataset show that the F1 value of this model reaches 67 .1 %, which is 5. 2% higher than the baseline model in this dataset, and the prediction of the overlap relation is also significantly improved. [ABSTRACT FROM AUTHOR]

Unlike long text classification, short text lacks information, and it is often difficult to obtain rich semantic features and extract complete syntactic features. Therefore, the effect of short text classification model needs to be improved . To solve this problem, this paper proposed a new model based on ResLCNN, the new model which was based on neural topic model and concatenated the features of multiple neural networks. Firstly, it extracted the topics to enrich the information of the short text. Secondly, it stored the topic information in the memory network and fused them with the sequence information as the final expression. Finally, it input the final expression into neural network with residual structure to extract local and global semantic features, and then combined the features together. The model achieves high accuracy in the Google Web search public dataset, which proves the effectiveness of the improved model in short text classification tasks. [ABSTRACT FROM AUTHOR]

How to allocate tasks efficiently and reasonably is one of the key problems in the research of spatial crowdsourcing. This paper proposed a MQTAM to solve the problem of low allocation efficiency of SC, and proposed an improved particle swarm optimization algorithm. The model took into account the time delay of task completion and the reliability of task completion, and used SCTAM_PSO algorithm to search the optimal assignment scheme to improve the task completion quality. The experimental results and analysis show that the MQTAM and SCTAM are effective and feasible. [ABSTRACT FROM AUTHOR]

In sound event detection tasks, the target event was susceptible to background noise, and wasn't present in a significantly high portion of sound signal flow. To solve the problem, this paper proposed a new method of sound event detection based on one-dimensional dense convolutional network (DenseNet ) with multi-level attention mechanism. Firstly, it used the one-dimensional DenseNet for frame-level detection, which was effective in finding the precise onset and offset time. Then, it introduced the multi-level attention mechanism in the one-dimensional DenseNet model, which made the attention-aware features from different modules change adaptively as layers went deeper. Therefore, the model could automatically select and attend on important frames for the targets while ignoring the unrelated parts (e. g. the background noise segments) . Finally, this paper evaluated the model using DCASE 2017 Task 2 development dataset. Results show that the overall performance of the proposed method has further improvement than the conventional deep learning method. [ABSTRACT FROM AUTHOR]

It is making great progress in the autonomous driving field with the rapid development of road scene understanding techniques. The safety is a concerning issue with respect to the real-time and accurate performance in the related tasks which contains the road segmentation, road classification and vehicle detection. To this end, this paper proposed an approach based on deep residual learning with an encoder-decoder network structure. On the one hand, the encoder network structure used different layers of residual networks to extract the abstract features in the high dimension, which shared in the next three tasks. On the other hand, the decoder network structure adopted a mechanism of parallel computing for sub-tasks, i. e., executed the road segmentation, vehicle detection and road classification tasks simultaneously. Additionally, it used the fully convolutional networks to upsample the extracted features to specifically solve the problem of road segmentation. At last, the experimental results show that the processing rate can effectively reach more than 15 fps with the high accuracy guaranteed. [ABSTRACT FROM AUTHOR]