引入注意力机制的纹样图像识别模型研究

Research on Pattern Image Recognition Model with Attention Mechanism

织物纹样的图像识别是纺织业中一项重要的技术,为提高民族织物纹样图像的准确识别率,保存纹样数字化信息,本研究以彝族织物为例,提出一种基于ResNet50模型与多头注意力机制相结合的纹样图像识别算法。首先使用扩充后符合训练要求的图片构建织物纹样数据库;然后构建ResNet模型,在模型中添加多头注意力机制模块,通过注意力机制获取的图像全局信息进行训练;最后采用GN(Group Normalization)层对样本通道求平均值,最后得出模型准确率为90.8%,较未加入注意力机制的ResNet模型提升了14.8%。基于此模型能快速准确识别复杂场景下拍摄或扫描的民族织物纹样图像,提高织物纹样识别的效率。

Image recognition of fabric patterns is an important technology in the textile industry.In order to improve the accurate recognition of ethnic fabric pattern images and preserve the digitized pattern information,a fabric pattern image recognition algorithm based on the combination of ResNet50 model and multihead attention mechanism was proposed in this study through the example of Yi fabrics.Firstly,a fabric pattern database was constructed using the images that have been expanded to meet the training requirements,and then the ResNet model was built with the multihead attention mechanism module.The global information of the image obtained by the attention mechanism was used for training.Finally,the GN(Group Normalization)layer was used to calculate the average value of the sample channels,and the model accuracy was acquired as 90.8%,which was 14.8%higher than the ResNet model without the attention mechanism.The method can quickly and accurately identify ethnic fabric pattern images captured or scanned in complex scenes,improving the efficiency of fabric pattern recognition.