<p>Defect and product number detection of highlight transparent objects (such as glass containers) are significant tasks in computer vision. However, existing highlight transparent object datasets have some problems: 1) lack large-scale and multi-modal data of actual scenes, 2) lack of efficient lightweight baseline. To address the above problems, this paper proposes the first large-scale visual-language glass container detection dataset (GCDet) for highlight transparent object defect and number detection. GCDet contains 36,500 images and 80,000+ instances, including subsets for defect detection and number detection. In addition, this paper customizes language templates for each object so that the dataset can be extended to visual-language tasks. Compared with the currently transparent object datasets such as Trans10 and GDD, GCDet has more diverse object categories and multi-modal visual-text annotations. GCDet provides valuable training data for multi-modal models in future industrial scenarios. Based on GCDet, this paper introduces a Visual-Language Knowledge Distillation (VLKD) architecture for highlight transparent object detection. VLKD introduces fine-grained language prompts into the highlight transparent object detection task, including three important innovations: distributed distillation, language prompt distillation, and decoupled distillation. Different from previous feature or object distillation, VLKD simultaneously considers the distribution relationship of features and the interaction of foreground and background to obtain a more refined feature transfer. In particular, VLKD introduces text into the distillation process for the first time, effectively improving the student model’s ability to understand key teacher features transfer. Extensive experiments on GCDet, MSCOCO, Trans10, Glass defect dataset, and Aluminum highlight defect dataset show that VLKD outperforms the current state-of-the-art methods, especially for detection of highlight transparent objects.</p>