FEEL: fast and effective emotion labeling, a dual ensemble approach for effective facial emotion recognition

Preprocessing

From the Kaggle platform, the standard datasets-FER2013 are obtained besides the IMED. The IMED dataset, which contains images representing blended emotions, was annotated using a semi-supervised strategy. Initially, candidate labels were proposed using automated prediction thresholds from baseline classifiers. These were then verified and refined through manual review by three independent annotators. Each image label was finalized through majority voting. This ensured that the emotional labels represent plausible combinations recognized consistently by human reviewers. The dataset is available at https://www.kaggle.com/datasets/msambare/fer2013. The former dataset contains images of common facial expressions, while the latter is for blended emotions. Prior to training, the datasets are preprocessed to have all the images in the same size and shape. With the os library, seven folders are created for the common facial emotions. The CSV file contents are retrieved, and the image files are saved in their respective folders after the mat arrays are made. Normalization at the pixel level is performed to avoid overfitting and ensure generalization.

Normalization

The technique of z-score normalization is utilized to bring the pixel values of all input images into a standard range of 0 and 1. For pursuing the process, mean $\mu$ and standard deviation $\sigma$ of the pixels of image $I$ are computed. A normalized image $I_N$ is obtained from an input image $I$ of channels ‘c’ with N pixels as,

(1) $$I_N={\displaystyle \frac{I_{x,y,c}-\mu }{\sigma }}$$where,

(2) $$\mu ={\displaystyle \frac{1}{N}\sum _{x,y,c}{I}_{x,y,c}}$$

(3) $$\sigma =\sqrt{{\displaystyle \frac{1}{N}\sum _{x,y,c}{\left(I_{x,y,c}-\mu \right)}^2}}.$$

In the equations, x and y represent the 2D positions of image pixels.

Image resizing

To bring all the images into the same standard size, images are resized. To preserve the pixels’ information while adjusting the dimensions, the scaling factor $S$ is ensured in every iteration. The factor $S$ for an input image $I$ with height $h$ and $w$ is given as, given as,

(4) $$S={\displaystyle \frac{h/w}{h^{\prime }/w^{\prime }}}$$where $h^{\prime }$, $w^{\prime }$ represents the height and width of the resultant image $I^{\prime }$

To retain the pixels’ intensity, the bicubic interpolation ( $B_I$) is applied, see Fig. 3. The operation is represented as,

(5) $$I^{\prime }=B_I\left(I,h^{\prime },w^{\prime }\right).$$

1^st phase ensemble-ESER

In the 1^st phase, the ESER model detects and recognizes basic emotions. The ensemble utilizes the three well-known CNN-based classifiers: DenseNet169, VGG16 and ReseNet50. After preprocessing, an unknown input image is passed through all the base models for prediction. As meta-features, the outputs of base models are fed to the meta-learner for precise recognition of emotion. The class of emotion for which the predicted accuracy score (PAS) of XGBoost is highest is supposed to be the class of input emotion. However, if the highest accuracy is less than 50%, the emotion is assumed to be blended. Therefore, the percent accuracies for sadness and happiness are stored in a separate CSV file-predicted mode in percent (PMP) for onward utilization in the second ensemble-ECER.

2^nd phase ensemble-ECER

The secondary statistical module (SSM) serves as an auxiliary decision-support component that integrates emotion likelihoods from ESER into the ECER prediction process. It receives class-wise probabilities of happiness and sadness from the first ensemble and recalibrates the ECER output using predefined emotion distribution weights. This hybrid scoring approach, calculated using Eq. (6), is particularly effective when blended emotions exhibit shared traits with basic emotional patterns. Detection and recognition of complex emotions at the second level are done using the ECER model. Suppose more than two distinct classes are predicted in the first level and the accuracy of meta learner is below 50% in the first level. This will indicate that the image carries a complex emotion and needs to be further classified. In the second level of the ensemble, the Xception, ViT and DenseNet221 are utilized to recognize the blended emotion. DenseNet221 was incorporated in ECER due to its extended depth and improved gradient flow across layers. Compared to shallower variants, DenseNet221 facilitates more refined feature extraction in deeper layers, which is particularly advantageous when dealing with nuanced patterns typical of blended emotional states. Its inclusion was based on empirical testing, where it showed superior feature abstraction for complex emotion patterns. To avoid the possibility of false recognition, the direct recognition by the ECER is complemented by supporting statistics in the SSM module. The percentage scores of happiness and sadness (obtained from the PMP) are used in SSM. The explicit formula used for the precise detection of the correct complex emotion is designed such that to give 60% weightage to direct recognition and 40% to the statistics behind the blended emotion. The net score (NS) accumulating the direct recognition (DR) score with the detected happiness (DH) and detected sadness (DS) is given in the Eq. (6).

(6) $$NS=\left(DR\ast {\displaystyle \frac{60}{100}}\right)+\left(DH\ast \left({\displaystyle \frac{S{H}_i}{100}}\right)+DS\ast \left({\displaystyle \frac{S{S}_i}{100}}\right)\right)\ast 40/100$$where $S{H}_i$ and $S{S}_i$ for $i=\left\{1,2,.,.,8\right\}$ represents the standard happiness and standard sadness scores, respectively as shown in Table 1. To enhance the reliability of blended emotion classification, a SSM has been incorporated. Its operation is presented formally in Algorithm 1. The module accepts emotion-specific confidence values—primarily happiness and sadness scores—from the first ensemble stage. These values are then adjusted using pre-defined statistical weights and combined with predictions from the second ensemble using Eq. (6). Based on the computed net score (NS), the final emotional state is inferred. The entire procedure has been designed to provide additional support in cases of emotional ambiguity.

Algorithm 1:

Algorithm of the proposed method.

DOI: 10.7717/peerj-cs.3138/table-6

The algorithm of the framework is shown in Algorithm 1.

Base models

In the first phase (ESER), DenseNet169, VGG16, and ResNet50 are used as base models, while in the second phase (ECER), Xception and ViT are utilized. The selection of base models was made based on their proven effectiveness in image-based tasks and complementary architectural characteristics. DenseNet169 was chosen for its dense connectivity and efficient gradient propagation; VGG16 for its simplicity and feature depth; ResNet50 for its residual mapping capabilities. In the ECER phase, Xception was employed due to its depthwise separable convolutions, which reduce parameters while maintaining performance, and ViT was selected for its strong contextual learning ability through self-attention. These models have consistently demonstrated competitive performance in visual recognition tasks and were found to offer diverse representational strengths. Details about the base models are presented in the following subsections.

Enhancing base model architecture

With each base model, the SE block is added to emphasize extracting informative features. The block is used to enhance the representational power of the neural networks. With the SE block, important features are adaptively weighted, thus allowing the base model to concentrate on the most relevant features. For an input tensor $T_m\in {\mathrm{R}}^{\mathrm{c}\times \mathrm{w}\times \mathrm{h}}$ with $c$ channels of height $h$ and width $w$, the block modulates the feature map via channel-wise scaling. Two main operations- squeezing and excitation are performed by the block. With squeezing, the spatial information is condensed into channels of a single value. For this purpose, global average pooling is performed to get vector. $V_z$,

(22) $$V_z=\left[v_1,v_2,v_3.,.,.,v_C\right]$$where $v_x$ being a global descriptor for channel $x$;

(23) $$v_x={\displaystyle \frac{1}{h\times w}\sum _{i=1}^h\sum _{j=1}^w{T}_{m\left(i,j,x\right)}.}$$With excitation, the importance of each channel is learned through the global optimizer $G_O\in R^C$. Two dense layers enact excitation. In between the two fully connected layers, non-linear activations is performed. The scaling factor $S_f$ used in the process is given as,

(24) $$S_f=\mathrm{s}\left({\mathrm{\omega }}_2\cdot \mathrm{\pi }\left({\mathrm{\omega }}_1{G}_O\right)\right).$$where ${\mathrm{\omega }}_1$ and ${\mathrm{\omega }}_2$ are the weights of first and second fully connected layers. $\mathrm{\pi }$ is the non-linear activation, while $\mathrm{s}$ is the sigmoid function. As shown in Fig. 4, the output of dense layers is received as input by the SE block. In the block, squeezing and excitation are performed, followed by output scaling. The output of the SE block is fed to the following fully connected layers; see Fig. 4.

Meta model

In ensemble learning, XGBoost is used as a meta-model. Outputs of the base models are fed to the XGBoost classifier to extract the complex meta-features relationships and overcome overfitting. The meta-model guarantees promising image-based classification and robust decision-making. XGBoost has a scalable decision tree architecture where each tree $t_x$ at iteration, $t$ is fit to the residual errors (gradients) of the previous trees. XGBoost was selected as the meta-learner owing to its strong regularization capabilities, high training speed, and superior performance in classification tasks involving high-dimensional and correlated features. While simpler methods such as logistic regression or more complex models like neural attention layers could have been employed, XGBoost was preferred due to its balance between computational efficiency and predictive accuracy in ensemble learning scenarios. Its gradient-boosted framework has been observed to effectively exploit inter-model output relationships without overfitting. In dataset D, if X represents the feature vector and Y is the target label, such that $X_i\in {\mathbb{R}}^d$ and $Y_i\in \mathbb{R}$

$\mathrm{D}={\left\{\left({\mathit{X}}_i,{\mathit{Y}}_{\mathit{i}}\right)\right\}}_{i=1}^n$, the objective function of XGBoost is given as,

(25) $$O_f=\sum _{i=1}^n{L}_s(Y_i,Y_i^p)+\sum _{j=1}^{T}R(t_j)$$where $L_s$ is the loss function, $Y_i^p$ is the predicted lable, $R$ is the regularization term, $t$ is for tree and T being the total number of trees. Regularization to avoid overfitting at tree ‘ $t_x$’ having $T_L$ tree leaves and weight $\omega$ is defined as follows.

(26) $$R\left(t_x\right)=\phi T_L+{\displaystyle \frac{1}{2}\sigma \sum _{i=1}^{T_L}{\omega }_i^2}$$where $\phi$ and $\sigma$ are the penalizing and regularization terms used in predicting label $Y_x^p$ which is defined as,

(27) $$Y_x^p=\sum _{i=1}^T{t}_i\left(X_i\right).$$The loss function $L_s$ of XGBoost is given as,

(28) $$L_s=\sum _{i=1}^n\left[\ell \left(Y_i,Y_i^{p\left(t-1\right)}\right)+G_{1_i}{t}_i\left(X_i\right)+{\displaystyle \frac{1}{2}{G}_{2_i}{t}_i{\left(X_i\right)}^2}\right]+R\left(t_i\right)$$where $l$ is the log loss, $G_{1_i}$ and $G_{2_i}$ are the first and second-order gradients, respectively.

Evaluation setup

A total of 3,750 images were taken from the Kaggle and IMED datasets for training, validation and testing. All the images were properly resized into fixed pixel sizes. Out of the total images, 70% were used for training, 15% for validation and 20% for testing. Before training, cross-validation was performed to avoid biases. In addition to cross-validation, several techniques were applied to prevent overfitting. Data augmentation methods such as horizontal flipping, random cropping, rotation, zooming, and brightness adjustments were consistently used across all training stages. Early stopping was also implemented during training to halt model updates if validation performance plateaued. Dropout layers were embedded within the classifier heads of certain base models to further enhance regularization. Stratified five-fold cross-validation was applied consistently across all models to maintain robustness and avoid bias. Furthermore, data augmentation techniques—including horizontal flipping, random rotations, zooming, and brightness variation—were uniformly performed during training. These steps were adopted to increase variability within the training set and reduce the likelihood of overfitting.

The results obtained during the two evaluations are locally stored in a separate text file for onward visualization and analysis. Following the establishment of the model architecture and training pipeline, experimental evaluation was conducted. The results and performance analysis of the ESER and ECER modules are discussed in the next section. An illustration of the recognition scores in the bar chart against the corresponding input images is presented in Fig. 5.

Evaluation of ESER

The performance of the first ensemble framework- ESER, is assessed along with the individual performance of the base models. The average accuracy of base models of ESER is 89%, overtaken by the ensemble as 95% accuracy is obtained for XGBoost. The average F1-score of the base models is 0.85; with the ensemble, the score is raised to 0.95. In order to validate the improvements observed, paired t-tests were conducted between the ensemble predictions and those of each base model. For both ESER and ECER stages, p-values below 0.01 were recorded, indicating that the accuracy improvements achieved by the proposed ensembles are statistically significant and unlikely to have occurred by chance.

The Area Under Curve (AUC), precision, recall, accuracy, and F1-score of the base models and the ensemble method using XGBoost are shown in Table 2. The formulas used for computing precision (P), recall (R) and F1-score are given in Eqs. (28), (29) and (30), respectively.

(29) $$P={\displaystyle \frac{TP}{TP+FP}}$$

(30) $$R={\displaystyle \frac{TP}{TP+FN}}$$

(31) $$F1=2\times {\displaystyle \frac{P\times R}{P+R}.}$$If $\alpha$ represents the number of correct classifications and $\beta$ the total number of predicted classes, accuracy A is given as,

(32) $$A={\displaystyle \frac{\alpha }{\beta }.}$$Similarly, if $p$ and $q$ are the coordinates of the points on the curve, then AUC is represented as,

(33) $$AUC\approx \sum _{j=1}^{j-1}{\displaystyle \frac{(p_{j+1}-p_j)}{2}\times \left(q_j+q_{j+1}\right).}$$As clear from Fig. 6, the rapid climb of the meta-model at the top-left indicates the effective performance of the model. Moreover, the positive rate at the top-left of XGBoost with the highest accuracy of 95% suggests that the proposed ensemble method optimally recognized the emotions.

The confusion matrix of the recognition of basic emotions is shown in Fig. 7, while the statistics are in Table 3.

Evaluation of ECER

To mitigate class imbalance within the blended emotion dataset (IMED), oversampling techniques were applied during training. Specifically, minority emotion classes were augmented via data replication and perturbation-based synthesis to ensure balanced class representation. Additionally, class weights were adjusted during loss calculation in ECER to penalize misclassification of underrepresented categories more heavily. Blended emotion recognition is assessed separately by testing the system with 200 images of complex facial emotions. Though the individual performance of the ViT model is satisfactory, the meta-model outperforms the individual accuracies of the base models. Without the ensemble approach, the average accuracy is 70.45%, whereas an acceptable accuracy of 88.7% is achieved with the ensemble. The precision, recall and F1-score of the base models and XGboost are presented in Table 4.

The reported average F1-score of the base models is 68% and 71%. However, a better F1-score (88%) was obtained for the ensemble. In addition to accuracy, other performance indicators such as F1-score, precision, recall, and ROC-AUC have been evaluated and presented in Tables S2A and S4A. These metrics were calculated to provide a more comprehensive understanding of the model’s classification behavior across all emotion categories. This implies that the ECER model effectively captures complex patterns and significantly lowers the chances of false negativity, as shown in Fig. 8.

From the first few epochs, it can be assumed that the model’s accuracy improves as the number of epochs is increased, see Fig. 9.

Pruning approach assessment

Though the performance of the ensemble model is superior to the base model in terms of accuracy, the aggregation performed by XGBoost adheres to additional computational cost. To reduce computational complexity, the pruning approach is assessed to analyze its effect regarding accuracy and computational cost. The approach intends to exclude the least significant base models during inferencing stage without affecting the performance as a whole. The two least significant models- DensNet from the ESER and Xception from the ECER, were excluded from the list of base models. With this, the accuracies obtained for ESER and ECER are 87% and 83%, respectively. This shows that about 21% of the computational time can be minimized at the cost of about a 7.05% reduction in the accuracies. Results of the ensemble obtained after the pruning approach are presented in Table 5.