【摘要】 针对卷积神经网络（CNN）多层卷积和池化操作导致的空间信息处理不足和全局上下文信息缺乏的问题，提出了一种基于CNN与Transformer的多尺度密集堆垛纸箱检测方法。设计特征提取融合模块，结合窗口自注意力机制，增强模型对全局特征的建模能力；引入跨尺度连接，融合更多不同层级的语义信息，使模型具备更大的感受野和更好的特征融合能力；提出BoxIoU损失函数用于边界框回归，通过计算边界框的最小点距和宽高比评估边界框相似性，提高模型的检测精度。实验结果表明：在密集堆垛纸箱数据集（SCD）上，该方法的mAP50达到了99.36%,mAP50-95达到了95.09%，具有良好的检测性能以及泛化能力。更多还原

【Abstract】 In response to the problems of insufficient spatial information processing and lack of global contextual information caused by multi-layer convolution and pooling operations in convolutional neural networks(CNN), a multi-scale densely stacked carton detection method based on CNN and Transformer is proposed. A feature extraction and fusion module is designed, combined with a window self-attention mechanism, to enhance the model′s ability to model global features. Cross-scale connections are introduced to fuse more semantic information at different levels, and the model has a larger receptive field and better feature fusion ability. A BoxIoU loss function is proposed for bounding box regression, which evaluates the similarity of bounding boxes by calculating the minimum point distance and aspect ratio of bounding boxes, improving the accuracy of the model. Experimental results show that on the SCD dataset of densely stacked cartons, the method achieves an mAP50 of 99.36% and an mAP50-95 of 95.09%, with good detection performance and generalization ability.更多还原