计算机技术与发展

对于织物疵点检测通常耗时、背景复杂、疵点种类多样且依赖人工操作的问题，提出了一种基于改进YOLOv7算法的轻量级检测方法。首先，增加了一个检测头Swin-Transformer,增强模型捕获和识别小目标特征的能力；其次，在主干特征提取阶段加入卷积注意力融合的注意力机制ACmix,使得模型可以自动学习并集中注意力在与小目标相关的区域或特征上，增加对重要特征信息的捕捉；采用自适应空间特征融合的方式，增强了多尺度特征融合能力；最后，使用Wasstertein距离优化损失函数，降低了模型对小目标的敏感度。通过对构建的含有6种疵点的面料数据集进行测试可以看出，模型ASFF-YOLOv7的检测精度mAP达到94%、帧率达到50.92帧/s,与其他8种算法相比，ASFF-YOLOv7的综合性能最优。

To address the issues commonly associated with fabric defect detection, such as time-consuming processes, complex backgrounds, diverse types of defects, and reliance on manual operations, we propose a lightweight detection method based on an improved YOLOv7 algorithm. Initially, a detection head, Swin-Transformer, is added to enhance the model's ability to capture and recognize features of small objects. Subsequently, a convolutional attention fusion mechanism, ACmix, is integrated during the backbone feature extraction phase, enabling the model to autonomously learn and focus on areas or features related to small objects, thereby enhancing the capture of crucial feature information. Adaptive spatial feature fusion is employed to bolster multi-scale feature integration. Finally, the Wasserstein distance is used to optimize loss function and reduce the sensitivity to small objects. Testing on a constructed fabric dataset containing six types of defects demonstrates that the ASFF-YOLOv7 achieves the highest detection accuracy and the fastest speed next to YOLOv7 in the dataset. Compared to the original algorithm, the proposed method achieves an accuracy rate of 94% and a frame rate of 50.92 frames per second.