Sarcasm is a complex linguistic phenomenon that often involves conveying sentiments contrary to the literal meaning of a statement. While traditional sarcasm detection models rely primarily on textual features, the increasing use of emojis in digital communication has introduced new challenges and opportunities. This paper explores the impact of integrating emoji semantics into sarcasm detection using deep learning. We benchmark four models—Bi-LSTM, CNN-LSTM, RoBERTa, and a novel Cross-Modal fusion model combining MiniLM and Emoji2Vec—on the emoji-annotated SarcOji dataset. Our findings show that the Cross-Modal model effectively captures sentiment incongruities and improves sarcasm detection, especially when textual and visual signals conflict. The study highlights the benefits of multimodal learning and underscores the necessity of modeling non-verbal cues like emojis to better understand sarcasm in modern digital communication.