The rapid increase in multimedia streams has created a greater demand for accurate and efficient detection of toxic content. Convolutional Neural Networks (CNNs) have traditionally performed well in this area due to their computational efficiency. However, they face challenges when dealing with toxic content that requires contextual understanding. In contrast, Vision-Language Models (VLMs) demonstrate strong semantic understanding, but their high computational costs limit their use for large-scale toxic content detection. To combine the efficiency of CNNs with the semantic understanding capabilities of VLMs, we propose BTCD, which uses CNNs for initial detection and then employs VLMs to correct misclassified samples from CNNs. Specifically, BTCD uses LogitNorm loss to enable CNNs to assign lower confidence scores to misclassified samples, allowing effective separation of incorrect samples through thresholding. Additionally, BTCD further enhances VLM recall through optimized prompts. Experiments on four datasets show that BTCD achieves the following advantages: compared to CNNs, BTCD improves accuracy by 5.1%–21.0%; compared to VLMs, BTCD achieves up to 1.8% higher accuracy while requiring only 32.1% of VLM inference time. Furthermore, our ablation studies explore the functionality of each component.