• ETRI Journal
• /
• v.42 no.1
• /
• pp.67-77
• /
• 2020

Image captioning has received significant interest in recent years, and notable results have been achieved. Most previous approaches have focused on generating visual descriptions from images, whereas a few approaches have exploited visual descriptions for image classification. This study demonstrates that a good performance can be achieved for both description generation and image classification through an end-to-end joint learning approach with a loss function, which encourages each task to reach a consensus. When given images and visual descriptions, the proposed model learns a multimodal intermediate embedding, which can represent both the textual and visual characteristics of an object. The performance can be improved for both tasks by sharing the multimodal embedding. Through a novel loss function based on class activation mapping, which localizes the discriminative image region of a model, we achieve a higher score when the captioning and classification model reaches a consensus on the key parts of the object. Using the proposed model, we established a substantially improved performance for each task on the UCSD Birds and Oxford Flowers datasets.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.416-418
• /
• 2022

Glioma is a type of brain tumor that occurs in glial cells and is classified into two types: high hrade hlioma with a poor prognosis and low grade glioma. Magnetic resonance imaging (MRI) as a non-invasive method is widely used in glioma diagnosis research. Studies to obtain complementary information by combining multiple modalities to overcome the incomplete information limitation of single modality are being conducted. In this study, we developed a 3D CNN-based model that applied input-level fusion to MRI of four modalities (T1, T1Gd, T2, T2-FLAIR). The trained model showed classification performance of 0.8926 accuracy, 0.9688 sensitivity, 0.6400 specificity, and 0.9467 AUC on the validation data. Through this, it was confirmed that the grade of glioma was effectively classified by learning the internal relationship between various modalities.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of Information Processing Systems
• /
• v.17 no.4
• /
• pp.754-771
• /
• 2021

In the task of continuous dimension emotion recognition, the parts that highlight the emotional expression are not the same in each mode, and the influences of different modes on the emotional state is also different. Therefore, this paper studies the fusion of the two most important modes in emotional recognition (voice and visual expression), and proposes a two-mode dual-modal emotion recognition method combined with the attention mechanism of the improved AlexNet network. After a simple preprocessing of the audio signal and the video signal, respectively, the first step is to use the prior knowledge to realize the extraction of audio characteristics. Then, facial expression features are extracted by the improved AlexNet network. Finally, the multimodal attention mechanism is used to fuse facial expression features and audio features, and the improved loss function is used to optimize the modal missing problem, so as to improve the robustness of the model and the performance of emotion recognition. The experimental results show that the concordance coefficient of the proposed model in the two dimensions of arousal and valence (concordance correlation coefficient) were 0.729 and 0.718, respectively, which are superior to several comparative algorithms.

• Journal of Intelligence and Information Systems
• /
• v.28 no.2
• /
• pp.101-125
• /
• 2022

Recently, with the development of computing technology and the improvement of the cloud environment, deep learning technology has developed, and attempts to apply deep learning to various fields are increasing. A typical example is anomaly detection, which is a technique for identifying values or patterns that deviate from normal data. Among the representative types of anomaly detection, it is very difficult to detect a contextual anomaly that requires understanding of the overall situation. In general, detection of anomalies in image data is performed using a pre-trained model trained on large data. However, since this pre-trained model was created by focusing on object classification of images, there is a limit to be applied to anomaly detection that needs to understand complex situations created by various objects. Therefore, in this study, we newly propose a two-step pre-trained model for detecting abnormal situation. Our methodology performs additional learning from image captioning to understand not only mere objects but also the complicated situation created by them. Specifically, the proposed methodology transfers knowledge of the pre-trained model that has learned object classification with ImageNet data to the image captioning model, and uses the caption that describes the situation represented by the image. Afterwards, the weight obtained by learning the situational characteristics through images and captions is extracted and fine-tuning is performed to generate an anomaly detection model. To evaluate the performance of the proposed methodology, an anomaly detection experiment was performed on 400 situational images and the experimental results showed that the proposed methodology was superior in terms of anomaly detection accuracy and F1-score compared to the existing traditional pre-trained model.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.17 no.6
• /
• pp.347-353
• /
• 2022

In this paper, we propose a method that performs a text-video retrieval model by replacing video properties using captions. In general, the exisiting embedding-based models consist of both joint embedding space construction and the CNN-based video encoding process, which requires a lot of computation in the training as well as the inference process. To overcome this problem, we introduce a video-captioning module to replace the visual property of video with captions generated by the video-captioning module. To be specific, we adopt the caption generator that converts candidate videos into captions in the inference process, thereby enabling direct comparison between the text given as a query and candidate videos without joint embedding space. Through the experiment, the proposed model successfully reduces the amount of computation and inference time by skipping the visual processing process and joint embedding space construction on two benchmark dataset, MSR-VTT and VATEX.

• Research in Plant Disease
• /
• v.30 no.1
• /
• pp.99-102
• /
• 2024

Integrated pest management is essential for controlling plant diseases that reduce crop yields. Rapid diagnosis is crucial for effective management in the event of an outbreak to identify the cause and minimize damage. Diagnosis methods range from indirect visual observation, which can be subjective and inaccurate, to machine learning and deep learning predictions that may suffer from biased data. Direct molecular-based methods, while accurate, are complex and time-consuming. However, the development of large multimodal models, like GPT-4, combines image recognition with natural language processing for more accurate diagnostic information. This study introduces GPT-4-based system for diagnosing plant diseases utilizing a detailed knowledge base with 1,420 host plants, 2,462 pathogens, and 37,467 pesticide instances from the official plant disease and pesticide registries of Korea. The AI plant doctor offers interactive advice on diagnosis, control methods, and pesticide use for diseases in Korea and is accessible at https://pdoc.scnu.ac.kr/.

• KIPS Transactions on Software and Data Engineering
• /
• v.6 no.4
• /
• pp.203-210
• /
• 2017

In this paper, we propose an effective neural network model for image caption generation and model transfer. This model is a kind of multi-modal recurrent neural network models. It consists of five distinct layers: a convolution neural network layer for extracting visual information from images, an embedding layer for converting each word into a low dimensional feature, a recurrent neural network layer for learning caption sentence structure, and a multi-modal layer for combining visual and language information. In this model, the recurrent neural network layer is constructed by LSTM units, which are well known to be effective for learning and transferring sequence patterns. Moreover, this model has a unique structure in which the output of the convolution neural network layer is linked not only to the input of the initial state of the recurrent neural network layer but also to the input of the multimodal layer, in order to make use of visual information extracted from the image at each recurrent step for generating the corresponding textual caption. Through various comparative experiments using open data sets such as Flickr8k, Flickr30k, and MSCOCO, we demonstrated the proposed multimodal recurrent neural network model has high performance in terms of caption accuracy and model transfer effect.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.10
• /
• pp.3955-3971
• /
• 2020

Due to the significant increase in the use of autonomous car technology, it is essential to integrate this technology with high-precision digital map data containing more precise and accurate roadway information, as compared to existing conventional map resources, to ensure the safety of self-driving operations. While existing map technologies may assist vehicles in identifying their locations via Global Positioning System, it is however difficult to update the environmental changes of roadways in these maps. Roadway vision algorithms can be useful for building autonomous vehicles that can avoid accidents and detect real-time location changes. We incorporate a hybrid architectural design that combines unsupervised classification of vision data with supervised joint fusion classification to achieve a better noise-resistant algorithm. We identify, via a deep learning approach, an intelligent hybrid fusion algorithm for fusing multimodal vision feature data for roadway classifications and characterize its improvement in accuracy over unsupervised identifications using image processing and supervised vision classifiers. We analyzed over 93,000 vision frame data collected from a test vehicle in real roadways. The performance indicators of the proposed hybrid fusion algorithm are successfully evaluated for the generation of roadway digital maps for autonomous vehicles, with a recall of 0.94, precision of 0.96, and accuracy of 0.92.

• KIISE Transactions on Computing Practices
• /
• v.22 no.11
• /
• pp.619-624
• /
• 2016

Social-aware video displays not only the relationships between characters but also diverse information on topics such as economics, politics and culture as a story unfolds. Particularly, the speaking habits and behavioral patterns of people in different situations are very important for the analysis of social relationships. However, when dealing with this dynamic multi-modal data, it is difficult for a computer to analyze the drama data effectively. To solve this problem, previous studies employed the deep concept hierarchy (DCH) model to automatically construct and analyze social networks in a TV drama. Nevertheless, since location knowledge was not included, they can only analyze the social network as a whole in stories. In this research, we include location knowledge and analyze the social relations in different locations. We adopt data from approximately 4400 minutes of a TV drama Friends as our dataset. We process face recognition on the characters by using a convolutional- recursive neural networks model and utilize a bag of features model to classify scenes. Then, in different scenes, we establish the social network between the characters by using a deep concept hierarchy model and analyze the change in the social network while the stories unfold.