This paper introduces a novel image-level ensemble learning strategy called Random Color Erasing (RCE) to enhance the robustness of deep learning models in scenarios with color variations. The method selectively erases partial or complete color information in training data without disrupting the original image structure, achieving a balanced weighting of color features and other features within the neural network. This approach mitigates overfitting and improves the model's ability to handle color variation, thereby enhancing its overall robustness. The proposed strategy is characterized by robust interpretability and has been tested across various tasks such as person re-identification and semantic segmentation. The results show that RCE significantly improves performance in cross-domain scenarios compared to existing methods that prioritize color robustness. The method is effective in reducing the model's reliance on color information, which is often the most salient and easily distinguishable feature, and enhances the model's ability to generalize across different domains. The experiments demonstrate that RCE outperforms baseline methods in terms of accuracy and robustness, particularly in cross-domain testing. The method is also effective in semantic segmentation tasks, where it improves performance on industrial dust and smoke segmentation. The paper also provides an analysis of the relationship between RCE and the generalization ability of neural networks, revealing the intrinsic reasons why networks trained with RCE may outperform ordinary networks. The main contributions of this paper include the introduction of RCE, an analysis of the network's performance with RCE from the perspective of classification, and extensive experiments on two distinct visual tasks and five baseline models with diverse architectures, demonstrating the effectiveness of the proposed ensemble learning strategy. The results show that the proposed method has significant potential to surpass traditional approaches in cross-domain testing.This paper introduces a novel image-level ensemble learning strategy called Random Color Erasing (RCE) to enhance the robustness of deep learning models in scenarios with color variations. The method selectively erases partial or complete color information in training data without disrupting the original image structure, achieving a balanced weighting of color features and other features within the neural network. This approach mitigates overfitting and improves the model's ability to handle color variation, thereby enhancing its overall robustness. The proposed strategy is characterized by robust interpretability and has been tested across various tasks such as person re-identification and semantic segmentation. The results show that RCE significantly improves performance in cross-domain scenarios compared to existing methods that prioritize color robustness. The method is effective in reducing the model's reliance on color information, which is often the most salient and easily distinguishable feature, and enhances the model's ability to generalize across different domains. The experiments demonstrate that RCE outperforms baseline methods in terms of accuracy and robustness, particularly in cross-domain testing. The method is also effective in semantic segmentation tasks, where it improves performance on industrial dust and smoke segmentation. The paper also provides an analysis of the relationship between RCE and the generalization ability of neural networks, revealing the intrinsic reasons why networks trained with RCE may outperform ordinary networks. The main contributions of this paper include the introduction of RCE, an analysis of the network's performance with RCE from the perspective of classification, and extensive experiments on two distinct visual tasks and five baseline models with diverse architectures, demonstrating the effectiveness of the proposed ensemble learning strategy. The results show that the proposed method has significant potential to surpass traditional approaches in cross-domain testing.