Existing methods for image emotion classification mainly rely on building large-scale networks that map visual features to emotion labels. However, as these networks grow in scale, they increasingly demand computational resources. Therefore, we propose a lightweight image emotion classification approach based on Chain of Thought (CoT) prompting. CoT introduces a novel paradigm in machine learning, leveraging specialized prompts to guide language models toward achieving appropriate outcomes. This paradigm has demonstrated promising few- shot performance across various natural language processing tasks but remains unexplored in the context of visual emotion understanding. Specifically, we first utilize Large Multimodal Models (LMMs) to generate reasoning processes that explain the emotional content in images. Then, we introduce a semantic similarity sampling mechanism to select the most relevant emotional reasoning examples for a given test sample, constructing appropriate CoT prompts to guide the LMM in step-by-step reasoning about emotional causes. Finally, we employ a self-consistency strategy to enhance the robustness of the reasoning process. Experimental results show that our approach outperforms mainstream methods on three widely used datasets, achieving superior accuracy in image emotion classification.