A cascaded spatiotemporal attention network for dynamic facial expression recognition

Ye, Yaoguang; Pan, Yongqi; Liang, Yan; Pan, Jiahui

doi:10.1007/s10489-022-03781-0

Cited by 3 publications

(2 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The model also gives the best results for the video inputs. A cascaded spatiotemporal attention network (CSTAN) is proposed [22] to integrate spatial and temporal emotional information. CSTAN helps in locating exact regions of interest for the given images.…”

Section: Literature Reviewmentioning

confidence: 99%

An Integrated CNN-BiLSTM Approach for Facial Expressions

Pansambal,

Nandgaokar,

Rajput

et al. 2024

IJACSA

View full text Add to dashboard Cite

Deep learning algorithms have demonstrated good performance in many sectors and applications. Facial expression recognition (FER) is recognizing the emotions through images. FER is an integral part of many applications. With the help of the CNN-BiLSTM integrated approach, higher accuracy can be achieved in identification of the facial expressions. Convolutional neural networks (CNN) consist of a Conv2D layer, dividing the given images into batches, performing normalization and if required flattening the data i.e. converting the data in a 1D array and achieving a higher accuracy. BiLSTM works on two LSTMs i.e. one in the forward direction and the other in a backward direction. One can use LSTM to process the images (datasets) however, it is suggested with the help of BiLSTM can predict the expressions with more accuracy. Input data is available in both the direction (forward and backward) which helps maintaining the context. Using LSTM CNN and BiLSTM always helps increasing the prediction accuracy. Application areas where a BiLSTM can give more prediction accuracy are the forecasting models, text recognition, speech recognition, classifying the large data and the proposed facial expression recognition. The integrated approach (CNN and BiLSTM) increases the accuracy significantly as discussed in the results and discussion section.This approach could be categorized as a fusion technique where two methods (approaches) are integrated to get higher accuracy. The results and discussion section elaborates the effectiveness of the integrated approach compared to HERO: human emotions recognition for realizing the intelligent internet of things. As compared to the HERO approach CNN-BiLSTM gives good results in terms of precision and recall.

show abstract

Section: Literature Reviewmentioning

confidence: 99%

An Integrated CNN-BiLSTM Approach for Facial Expressions

Pansambal,

Nandgaokar,

Rajput

et al. 2024

IJACSA

View full text Add to dashboard Cite

show abstract

“…Model processes the video as sequences of different frames and each frame consists of images. Model processes the features extracted from given images automatically by introducing 03 steps as follows - Authors [3,22] have introduced robots and embed a proposed model to process real-life images. In this model robot extracts images from camera, video etc.…”

Section: Literature Reviewmentioning

confidence: 99%

Integrating Dropout Regularization Technique at Different Layers to Improve the Performance of Neural Networks

Pansambal¹,

Nandgaokar²

2023

IJACSA

View full text Add to dashboard Cite

In many facial expression recognition models it is necessary to prevent overfitting to check no units (neurons) depend on each other. Therefore, dropout regularization can be applied to ignore few nodes randomly while processing the remaining neurons. Hence, dropout helps dealing with overfitting and predicts the desired results with more accuracy at different layers of the neural network like 'visible', 'hidden' and 'convolutional' layers. In neural networks there are layers like dense, fully connected, convolutional and recurrent (LSTM-long short term memory). It is possible to embed the dropout layer with any of these layers. Model drops the units randomly from the neural network, meaning model removes its connection from other units. Many researchers found dropout regularization a most powerful technique in machine learning and deep learning. Dropping few units (neurons) randomly and processing the remaining units can be considered in two phases like forward and backward pass (stages). Once the model drops few units randomly and select 'n' from the remaining units it is obvious that weight of the units could change during processing. It must be noted that updated weight doesn't reflect on the dropped units. Dropping and stepping-in few units seem to be very good process as those units which step-in will represent the network. It is assumed to have maximum chance for the stepped-in units to have less dependency and model gives better results with higher accuracy.

show abstract