DEEP LEARNING APPROACHES TO EMOTION RECOGNITION IN PHOTOGRAPHIC IMAGES

Photo Emotion Recognition (PER) is supposed to learn what emotion is expressed or invoked by an image based on visual representations of color harmony, composition, object-scene semantics, human expressions in the presence when possible. In contrast to face-centric affect analysis, PER needs to analyze the emotions that frequently are a result of situational semantics and aesthetics, as opposed to explicit facial expression. This enhances ambiguity, label subjectivity, and overlapping of the classes. Additionally, the benchmarks of PER are often characterized by class imbalance and noisy annotations because of the different human perceptions. The paper is a complete analytical PER study with a proposed hybrid deep learning model (combines convolutional representations and transformer) to simultaneously identify low-level aesthetic representations and global semantic context. The proposed architecture includes CNN and transformer branches with regard to local texture color stimuli and long-range relational reasoning respectively, followed by the gated-feature fusion and using a balanced classification head. Class-balanced focal loss, label smoothing and emotion-preserving augmentation are used to construct a robust training pipeline, which prevents the distortions that are likely to alter affective meaning. The assessments of the results include macro-F1, per-class sensitivity, and the confusion behavior among the neighbouring emotions, calibration, and cross-domain strength. Numerous experiments of ablation prove that fusion and high-resistance loss decisions are always more effective on the macro-F1 and assist less in common confusions (e.g., fear vs. surprise, sadness vs. contentment/neutral). Lastly, it is a case of explainability analysis through gradient-based localization to determine whether the predictions are in agreement with the emotionally salient regions. Conclusion of the paper is deployment advice (latency, model size, and quantization) and ethical inferences of subjective affect modelling.