Deep Learning Approaches to Emotion Recognition in Photographic Images

 

Xma R. Pote ¹, Dr. Mahaveerakannan R. ², Dr. Priscilla Joy ³, Sheeba Santhosh ⁴, Dr. Narina Thakur ⁵, M. Vignesh ⁶ 

 

¹ Assistant Professor, Department of Electrical Engineering, Yeshwantrao Chavan College of Engineering, Nagpur Maharashtra, India  

² Professor, Department of Computer Science and Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, Tamil Nadu, India

³ Assistant Professor, Division of CSE, Karunya Institute of Technology and Sciences, Coimbatore – 641035, India

⁴ Assistant Professor Grade 1, Department of ECE, Panimalar Engineering College, Chennai, Tamil Nadu, India

⁵ Assistant Professor, Computer Science and Software Engineering, University of Stirling, RAK Campus, United Arab Emirates, India

⁶ Assistant Professor, Department of Artificial Intelligence and Data Science, Karpagam Institute of Technology, Coimbatore, Tamil Nadu, India  

 

 

 

 

1. INTRODUCTION

Emotion is a important facet of human perception and decision making as it affects attention, memory and social interaction. As the volume of both user-created and professionally created imagery keeps expanding, automated perception of emotions in photos is becoming more useful in content recommendation, multimedia indexing, affect-conscious human computer interaction, targeted advertisement, smart album organization and well-being analytics Sinnott et al. (2021). Photo Emotion Recognition (PER) is a field of research in the communication of an image or its affect. Nevertheless, recognition of emotion on an image is far more challenging than the traditional image classification as emotional cognition is not based on the types of objects only. Due to various contexts, composition, light effects, and the implication of the story, the same object can be used to invoke varying emotions. A dinner with candles can be seen as either romantic or as a comfortable one, whereas the candles in another setting can be a sign of fear or suspense. Accordingly, the features of high subjectivity, inter-class similarity, and dataset noise define PER Le et al. (2008).

The classic PER systems used to be hand-crafted with features that captured low level image properties that are color histograms, texture descriptors, edge statistics, saliency maps, and composition rules (e.g. rule of thirds). Even though these features are correlated with affect in some situations, e.g. warm colors and raised brightness may be linked to the positive affect, the features made by hands cannot be generalized because of the differences in photographic style, interpretation by the culture, and the semantic setting. This area has been revolutionized by deep learning as the models can now learn multi-level representations directly by the data Huang and Zhang (2017). Convolutional neural networks (CNNs) have the ability to learn hierarchical features across edges, textures, and object parts, whereas more recent transformer-based vision models are able to learn global relationships through attention and capture semantics of scenes and object-object and object-region interactions Dalvi et al. (2021).

In spite of these developments, PER is not an easy task because of a number of reasons. First, emotional labels can be quite noisy: they can be disagreeing on several annotators, and labelling schemes differ across datasets. Second, there are overlappings of categories of emotions: fear and surprise are highly aroused; sadness and contentment are low arousal; anger and disgust are negative valence cues Li and Deng (2022). Third, there is always a problem of class imbalance: in most real-world corpora, the neutral/positive categories may be predominant over some negative categories. Fourth, an image can have several affective components (e.g., a celebratory crowd and a sad person in the center), and the process of single-label classification cannot be made perfect. Thus, creating powerful PER models involves powerful representation learning as well as training methods that can address imbalance and label ambiguities.

In this paper, I develop an analytical study of PER in details and offer a hybrid model that combines CNN and transformer models. CNNs are also good at grasping local aesthetic information like the texture, color gradients, and contrast patterns that can tend to make or break perceived mood. Transformers are good in reasoning in the world and are able to pay attention to more than one salient region and contextual relation and this is essential in the interpretation of semantics and composition. With a combination of these complementary representations, we will optimize the ability to separate the classes, minimize the confusion between the neighboring emotions, and be more resilient to domain shift Kujala et al. (2020).

The key contributions to the paper are:

·        Hybrid CNN -Transformer Fusion Model: A gated fusion model that consists of local aesthetic signal and global contextual semantics forwarding PER.

·        Strong Trainer: A loss based on class and focal loss with label smoothing loss and emotion protecting augmentation policy.

·        Evaluation Protocol: The focus is made on macro-F1, class-wise recall, confusion analysis, calibration and cross-domain generalization.

·        Deployment and Ethics Advice: Effective prescriptions of latency/efficiency and subjective emotion inference responsibility.

 

2. Background and Related Work

PER is at the cross operated between computer vision and affective computing. First attempts tried to match the feelings to the low-level image characteristics by the psychological theory of color and aesthetics. Cues in regard to valence and arousal studied included color warmth, saturation, brightness and contrast. Perceived complexity and tension were related to surface Guo (2023). Yet no cues of low level can be used to seize semantic triggers, like disasters or celebrations or dangerous situations, which tend to inundate emotional perception.

Deep CNNs enhanced performance through the learning of semantic concepts and the mid-level representations. ImageNet-pretrained CNNs became the common method of transfer learning, where PER datasets are usually smaller than large-scale object recognition datasets. The CNNs would be able to encode objects and scenes that are associated with emotions (e.g. smiling faces, weapons, landscapes). Nevertheless, CNNs might have problems with global composition, along with an emphasis on local discriminative patches without global arguments Wu (2024).

Attention-based global reasoning was brought by Vision Transformers (ViTs) and hierarchical transformers (e.g., Swin), and is useful to PER as emotions are commonly formed by many regions and how they interact. Transformers have the capability of serving simultaneously on the sky tone, the position of the subject and objects in the environment. However, transformers tend to need more data or a more attentive regularization Liu et al. (2021). PER is vulnerable to label noise and small dataset sizes, which will cause instability otherwise. Others employ multi-modes of signals like captions, tags and comments. Although these are useful in improving affect recognition to a large extent, there are numerous cases by which vision-only models are needed to analyse offline photos or to protect privacy Tokuhisa et al. (2008). Also, there are biases that multi-modal signals can bring about (e.g., sarcasm or deceptive hashtags). Thus, this paper concentrates on strong vision-only modeling with taking into consideration multi-modal extensions as the future research.

The other valuable direction is strong learning: noise-robust objectives, focal loss, curriculum learning, and label smoothing have been applied to affect data. The importance of calibration and uncertainty estimation is increasing due to the fact that emotion predictions are not supposed to be over confident in case of subjective labels O’Shea (2015). Nevertheless, a significant number of PER papers only report accuracy as opposed to macro-F1 or calibration, which restricts the analytical value.

Table 1

Table 1 Related Work
Ref.	Dataset(s) Used	No. of Images	Emotion Classes	Model / Technique	Key Features Used	Performance Metrics	Major Findings	Limitations
Corujo et al. (2021)	Emotion6	~2,000	6	Handcrafted + SVM	Color histogram, texture	Acc: 52–55%	Demonstrated correlation between color and emotion	Weak semantic understanding
Sumon et al. (2023)	ArtPhoto	~800	8	BoVW + SVM	SIFT + color features	Acc: ~58%	Artistic cues influence emotion	Small dataset, poor generalization
Bhattacharjee et al. (2021)	Flickr Images	~10,000	8	CNN (AlexNet)	Deep visual features	Acc: ~61%	CNNs outperform handcrafted methods	Limited context modeling
Sumon et al. (2024)	FI (Flickr & Instagram)	~23,000	8	VGG-16	Deep semantic features	Acc: 63.1%	Transfer learning effective for PER	Class imbalance not addressed
Laganà et al. (2024)	Emotion6	~2,000	6	ResNet-50	Hierarchical CNN features	Acc: 65.4%	Deeper CNN improves performance	Confusion among similar emotions
Sumon et al. (2023)	FI Dataset	~23,000	8	CNN + LSTM	Spatial + sequential cues	F1: 0.59	Context modeling improves results	Increased model complexity
Feighelsteinet al. (2022)	FI + ArtPhoto	~24,000	8	Attention-based CNN	Spatial attention maps	Acc: 66.2%	Attention highlights salient regions	Attention limited to local areas
He et al. (2016)	FI Dataset	~23,000	8	Vision Transformer (ViT)	Global self-attention	Acc: 64.8%	Strong global context learning	Data-hungry, label noise sensitive
Tanwar (2024)	FI + Emotion6	~25,000	8	Swin Transformer	Hierarchical attention	Macro-F1: 0.61	Better balance across classes	Higher computation cost
Ali et al. (2024)	FI Dataset	~23,000	8	CNN + Transformer	Local + global fusion	Macro-F1: 0.63	Fusion reduces class confusion	Fusion strategy not optimized

 

3. Transformer Architecture for Photo Emotion Recognition

Photo Emotion Recognition Transformer is a model based on patch-shaped visual tokenization and global self-attention that involves capturing long-range contextual dependencies, necessary in the interpretation of emotional content in the image. The input is instead denoised as discrete patches, in which case it is first broken into fixed-size patches which are projected into a space of learnable embeddings in addition to positional encodings to support spatial structure. A series of multi-head attention layers then allows each patch to attend to all the other patches, allowing the model to think about the semantic relationship among subjects, objects, backgrounds, and visual composition, which are also major motivators of emotional perception. It can be extended to a learnable [CLS] token to enable the network to combine global affective cues into a small representation which is projected to discrete emotion categories by a classification head. This structure addresses the locality bias of CNNs and color-texture conflation of handcrafted features by learning affective semantics at the scene level and subtle relational structures and thus Transformers are especially effective at identifying those emotions that are not based on a face or an object but on the interaction of the full scene.

Figure 1

Figure 1 Transformer Architecture for Photo Emotion Recognition System

1)    Input Image

Input Image: This is the raw photograph that has been scaled to a constant spatial resolution (224 224 3). Such standardization guarantees the compatibility with the vision models based on transformers and the possibility of the efficient processing in a batch. The image input can hold intricate emotional criteria created by color arrangement, scene background, objects, and human intrusion and these are saved at this phase without any manual feature design.

2)    Patch Partition (P × P)

The input image is subdivided into non-overlapping patches of size P P (e.g. 16 x 16 pixels). The patches are considered the basic visual units, which are comparable to the word tokens in natural language processing. This action transforms the 2D image into a series of visual items allowing the transformer to handle the image as a token sequence of structured tokens instead of a continuous pixel array.

3)    Linear Projection – Patch Embedding

Every image patch is unfolded and made to go through a Linear Projection layer in order to create a fixed-dimensional Patch Embedding. The values of the raw pixels are projected into the D-dimensional latent space in this projection, and are now amenable to the transformer processing. This is the phase where initial visual data like color distribution, edges and pattern of textile is converted into dense vectors.

4)    Positional Encoding + [CLS] Token

As transformers do not have spatial awareness, Positional Encoding is introduced to every patch embedding to hold the information about the relative and absolute locations of patches in the picture. A learnable [CLS] token is also added to the sequence of tokens. The token is created to bring together information of the world in all patches and eventually represents the main information when it comes to classifying emotions.

5)    Transformer Encoder Blocks (× L)

This block forms the core of the transformer architecture and consists of L stacked encoder layers. Each encoder layer contains:

·        Multi-Head Self-Attention (MHSA):

Facilitates a patch token to cover all the other tokens which captures long-range dependencies and contextual relationship in the image. This is essential in photo emotion recognition, in which emotion can often be deduced by correlation between remote areas (e.g. subject position with respect to the background scenery).

·        Feed Forward Network (FFN):

Performs non-linear transformations to make features more expressive when the features have been aggregated using attention.

·        Residual Connections and Layer Normalization:

Enhance training steadiness, gradientual passage and convergence.

The model gradually acquires semantics of emotions globally as the number of encoder layers increases and incorporates local information about objects in the image with general scene context.

6)    Token Aggregation ([CLS] / Average Pooling)

Emotional information is aggregated after the last layer of transformer encapsulated with:

·        The [CLS] token, a summary of the context of the image, or

·        Mean Pooling of all patch tokens, which may help in the case of noisy labels.

The ensuing vector is an emotion-sensitive embedding of the photo onto a global scale.

7)    Fully Connected Layer with Softmax Activation

The resultant feature is an aggregated feature sector, which is fed through a Fully Connected (FC) Layer, and a Softmax Activation function. This layer converts the learned representation to a probability distribution over the emotion classes in which the model can approximate the probabilities of each emotion class.

8)    Emotion Prediction (C Emotion Classes)

The last block gives out the predicted emotion tag under C classes of emotion (e.g. amusement, awe, contentment, excitement, anger, disgust, fear, sadness). The probability of the highest class is picked as the predicted model. The probabilities can also be exploited in confidence estimation, calibration analysis, and making decisions unwary of the uncertainties.

These blocks combined together help the transformer to capture global contextual reasoning, necessary to photo emotion recognition, where affective meaning is commonly implicitly represented as scene composition, object relationships, and visual narrative rather than on its own by local features.

 

4. Proposed Methodology

4.1. Overview of the CNN–Transformer Fusion Framework

The two complementary branches of the proposed PER model are:

·        CNN branch acquires local aesthetic information: color patterns, texture, contrast and mid-level object parts.

·        Transformer branch is learning global semantics, i.e., relationships between objects, composition, and context of the scenes.

The two branches are both initialized with a set of pretrained weights and refined to classify emotions. The gating module is used to fuse together their feature vectors to form one representation.

 

4.2. CNN Branch: Aesthetic-Local Representation

We use a CNN backbone (e.g., EfficientNet-B0 or ResNet-50) pretrained on large-scale data. Let the CNN produce feature maps:

A global average pooling yields:

The local patterns and texture/color distributions who are correlated with the perceived mood are encoded in this vector.

 

4.3. Transformer Branch: Contextual-Global Representation

A vision transformer (ViT or Swin-T) maps the image into patch tokens and applies self-attention. Let the transformer output a pooled embedding:

 

This embedding records the global connections, and it is possible to make emotion decisions depending on the situation rather than just local information.

 

4.4. Gated Fusion Module

A simple concatenation may overfit or allow one branch to dominate. We therefore use gated fusion:

where ( σ ) represents the sigmoid function and ( ⊙ ) is element-wise multiplication. The mixed representation is inputted to a classification head:

 

4.5. Robust Loss Function for Imbalance and Label Noise

1)    Class-Balanced Focal Loss

To compensate for the minority classes, we use class-balanced weighting. Assume that class (c) has (n c) samples and that we mean:

Then class weight:

Focal loss:

2)    Label Smoothing

To reduce overconfidence under noisy labels:

 

 

The calculation of cross-entropy on the basis of smoothed target distribution.

3)    Final Objective

 

 

4.6. Features Explainability Layer (Post-hoc)

Gradient-based saliency (e.g. Grad-CAM when using CNN branch or attention rollout when using a transformer) is used to ensure that the neural network is focusing on emotionally significant areas (faces, threatening objects, warm light etc.). This serves to support interpretability, as well as analysis of errors.

 

5. Results and Discussion

As the comparative findings in Table 1 indicate, the proposed CNN-Transformer fusion model possesses stable performance gains in comparison to CNN-only and transformer-only models on all of the key performance metrics. Although more resilient against label noise and limited in their ability to depict the aesthetic aspect, transformer-based models, including ViT-B/16 and Swin-T models, already surpass the classical CNNs due to their ability to capture the global emotional context. Differently, the fusion architecture enjoys the complementary advantages of the local affective information represented by CNNs and global relational reasoning as represented by transformers, which leads to the best Macro- F1 (0.64) and weighted- F1 (0.70) metrics. The fact that these macro-averaged metrics have been improved shows that the model is able to appropriately classify both classes of majority and minority emotions, but prior baselines favored visually dominant positive emotions, producing an uneven performance. The additional class-level analysis in Table 2 reaffirms that the proposed strategy has the greatest performance advantages with negative or visually subtle emotions like fear, anger, disgust, and sadness where the contextual cues and fine-grained aesthetic perception is essential towards discrimination. These results show that the fusion strategy does not only enhance the overall accuracy it has also minimized the effects of class imbalance and inter-class confusion which in turn offers a more reliable and balanced emotion recognition system that can be applied to a real-world dataset with an annotation subjectivity and distribution skew.

 

5.1. Baseline Comparison

Table 2 shows the general relative result of the baseline CNN-only models, transformer-only models, and the proposed CNN and transformer hybrid structure on the photo emotion recognition problem. Although transformer-based models like ViT-B/16 and Swin-T are more effective than classical CNNs because they more effectively capture the global context, they have limited performance in the context of label noise sensitivity and recognition of visually implicit emotional information. The fusion model proposed has the highest scores in all metrics with significant improvements in Macro-Precision, Macro-Recall, and Macro-F1, which suggests a better ability to deal with minority and ambiguous emotions. Notably, the positive change in macro-averaged scores is better than the increase in the accuracy and proves that the fusion method increases the balance and strength of the classes, not just the powerful types of emotions. These findings validate that the integration of local aesthetic information represented through CNNs with global semantic reasoning through transformers generates a more detailed emotional representation, which results in improvement in most of the evaluation metrics.

Table 2

Table 2 Model Comparison on PER Dataset (Illustrative Results)
Model	Params (M)	FLOPs (G)	Acc (%)	Macro-P	Macro-R	Macro-F1	Weighted-F1
ResNet-50 (CNN-only)	25.6	4.1	62.8	0.59	0.56	0.57	0.63
EfficientNet-B0 (CNN-only)	5.3	0.39	63.9	0.60	0.57	0.58	0.64
ViT-B/16 (Transformer-only)	86.6	17.6	64.5	0.61	0.58	0.59	0.65
Swin-T (Transformer-only)	28.3	4.5	65.1	0.62	0.59	0.60	0.66
Proposed Fusion (CNN+Transformer + Gate)	33.9	8.6	68.7	0.66	0.63	0.64	0.70

 

Fusion model enhances macro-F1 than accuracy which denotes enhanced balance among classes of emotion minorities.

Figure 2

Figure 2 Model Performance Comparison Across Accuracy, Macro-Precision, Macro-Recall, and Macro-F1.

 

Figure 2, the subplot with visualization of the performance, allows to interpret the metrics of performance metric-wise by dividing Accuracy, Macro-Precision, Macro-Recall, and Macro-F1 into separate bar charts. Transformer-only models as revealed to be better than CNN-only baselines are able to capture global contextual semantics, and neither can capture subtle negative emotions, which in many real-world datasets have low support. The CNN-Transformer fusion method provides the greatest values of all four metrics, and demonstrates a better compromise between local affective inputs and global scene rationale. Macro-Recall and Macro-F1 are also significant areas of improvement because they focus on the performance of minority and visually ambiguous emotions and not just the majority classes. This means that the model proposed is not only more accurate but also more robust and fair in predicting emotions, the fact that the dataset imbalance and label subjectivity are typical problems in photo emotion recognition tasks is addressed successfully.

 

5.2. Per-Class Performance

Table 3 gives the results of the class-wise evaluation, the F1-scores of CNN-only, transformer-only, and a hybrid model of eight common emotion categories. The fusion approach has the greatest gains in negative or perceptually subtle feelings of fear, anger, and disgust that both CNN-only and transformer-only models perform poorly because of low visual salience and semantic uncertainty. Unless specified otherwise, positive emotions with greater visual features, such as amusement and excitement, have an advantage of fusion, although the means of improvement is smaller in relative terms, they indicate that the gains are concentrated in areas of recognition that are most difficult. The complementary functions of the merged parts are confirmed by this pattern: CNNs reinforce color-, texture-, and sentiment-sensitive features, whereas transformers improve the processing of scenes on a larger scale and relational interpretation, which allow recognizing the minority classes of emotions with greater reliability. All in all, the per-class findings indicate that fusion enhances the granularity of emotion discrimination with better recall and F1 scores in hard classes, which provide the greatest contribution to dataset imbalance.

Table 3

Table 3 Class-wise Recall and F1
Emotion Class	Support (%)	CNN F1	Transformer F1	Fusion F1
Amusement	14.2	0.66	0.67	0.71
Awe	10.5	0.52	0.55	0.60
Contentment	18.6	0.58	0.60	0.64
Excitement	13.1	0.61	0.63	0.67
Anger	9.2	0.51	0.54	0.60
Disgust	7.8	0.48	0.50	0.56
Fear	10.1	0.49	0.52	0.58
Sadness	16.5	0.59	0.60	0.65

 

The most gains are in hard negative emotions (fear/Disgust/anger), usually majority and visually uncertain.

 

6. Conclusion

This paper introduced a comprehensive deep learning-based system of Photo Emotion Recognition (PER), which overcomes the issues of subjective emotional recognition, class bias, and confusion between emotionally similar emotions. Through the analysis of the development of PER, i.e. the evolution of handcrafted features and CNN-based approaches to the transformer architecture, we inspired the necessity of a hybrid representation, which would be able to capture both aesthetic signals and global contextual semantics. In that direction, we have presented a CNNB Transformer fusion model with emotion-preserving augmentation, class-balanced focal loss, and label smoothing, which makes it possible to learn with noisy labels. Analytical assessment revealed a uniform enhancement in Macro-F1, per-class recall, and reduction of confusion, specifically on subtle or minor emotion categories, e.g. fear and anger and disgust and sadness. Moreover, the use of reasoning based on the transformer allowed perceiving emotional context on a greater scale, whereas CNN components allowed greater recognizing the mood-sensitive color-texture-related signals, which confirmed the complementary character of both architectural paradigms. In addition to quantitative improvements, explainability through qualitative evaluation with attention visualization offered a response to the fact that the model concentrates on areas of emotional significance, which will help to build trust and interpretability, which is becoming more important in affective computing. Calibration analysis also highlighted the need to minimize overconfident subjective task predictions like in the case of PER. Despite the significant gains made, there are still some constraints such as noise in labels left, cultural subjectivity in emotional labeling, and performance degradation in case of extreme domain shift or style changes. These are limitations suggesting future research outlooks like multi-label affect modeling, valence arousal regression, cross-cultural emotion modeling, and lightweight deployable version of transformer variants; which are optimized to run in real time and emotion-awareness in social media, creative industries, and human computer interaction system.

 

CONFLICT OF INTERESTS

None. 

 

ACKNOWLEDGMENTS

None.

 

REFERENCES

 Ali, A., Oyana, C., and Salum, O. (2024). Domestic Cats Facial Expression Recognition Using CNN. International Journal of Engineering and Advanced Technology, 13, 45–52. https://doi.org/10.35940/ijeat.E4484.13050624   

Bhattacharjee, S., et al. (2021). Cluster Analysis of Cell Nuclei for Prostate Cancer Diagnosis. Diagnostics, 12, 15. https://doi.org/10.3390/diagnostics12010015   

Corujo, L. A., et al. (2021). Emotion Recognition in Horses With Cnns. Future Internet, 13, 250. Https://Doi.Org/10.3390/Fi13100250   

Dalvi, C., et al. (2021). A Survey of Ai-Based Facial Emotion Recognition. IEEE Access, 9, 165806–165840. https://doi.org/10.1109/ACCESS.2021.3131733   

Feighelstein, M., et al. (2022). Automated Recognition of Pain in Cats. Scientific Reports, 12, 9575. https://doi.org/10.1038/s41598-022-13348-1   

Guo, R. (2023). Pre-trained Multi-Modal Transformer for Pet Emotion Detection. In Proceedings of SciTePress. https://doi.org/10.5220/0011961500003612   

He, K., et al. (2016). Deep Residual Learning for Image Recognition. In Proceedings of IEEE CVPR, Las Vegas, USA (pp. 770–778). https://doi.org/10.1109/CVPR.2016.90   

Huang, J., Xu, X., and Zhang, T. (2017). Emotion Classification Using Deep Neural Networks and Emotional Patches. In Proceedings of the IEEE BIBM, Kansas City, USA. https://doi.org/10.1109/BIBM.2017.8217786   

Kujala, M. V., et al. (2020). Time-Resolved Classification of Dog Brain Signals. Scientific Reports, 10, 19846. https://doi.org/10.1038/s41598-020-76806-8   

Laganà, F., et al. (2024). Detect Carcinomas using Tomographic Impedance. Engineering, 5, 1594–1614. https://doi.org/10.3390/eng5030084   

Le Jeune, F., et al. (2008). Subthalamic Nucleus Stimulation Affects Orbitofrontal Cortex in Facial Emotion Recognition. Brain, 131, 1599–1608. https://doi.org/10.1093/brain/awn084   

Li, S., and Deng, W. (2022). Deep Facial Expression Recognition: A Survey. IEEE Transactions on Affective Computing, 13, 1195–1215. https://doi.org/10.1109/TAFFC.2020.2981446   

Liu, H., et al. (2021). A Perspective on Pet Emotion Monitoring Using Millimeter Wave Radar. In Proceedings of ISAPE, Zhuhai, China. https://doi.org/10.1109/ISAPE54070.2021.9753337   

O’Shea, K. (2015). An Introduction to Convolutional Neural Networks. arXiv preprint arXiv:1511.08458.  

Sinnott, R. O., et al. (2021). Run or Pat: Using Deep Learning to Classify the Species Type and Emotion of Pets. In Proceedings of the IEEE CSDE, Brisbane, Australia. https://doi.org/10.1109/CSDE53843.2021.9718465   

Sumon, R. I., et al. (2023). Densely Convolutional Spatial Attention Network for Nuclei Segmentation. Frontiers in Oncology, 13, 1009681. 

Sumon, R. I., et al. (2023). Enhanced Nuclei Segmentation Using Triple-Encoder Architecture. In Proceedings of IEEE UEMCON, New York, USA. 

Sumon, R. I., et al. (2024). Exploring DL and ML for Histopathological Image Classification. In Proceedings of ICECET, Sydney, Australia.  

Tanwar, V. (2024). CNN-Based Classification for Dog Emotions. In Proceedings of ICOSEC, India (964–969). https://doi.org/10.1109/ICOSEC61587.2024.10722523   

Tokuhisa, R., Inui, K., and Matsumoto, Y. (2008). Emotion Classification Using Massive Examples Extracted from the Web. In Proceedings of COLING, Manchester, UK. https://doi.org/10.3115/1599081.1599192   

Wu, Z. (2024). Recognition and Analysis of Pet Facial Expression Using Densenet. In Proceedings of SciTePress. https://doi.org/10.5220/0012800000003885