改进ResNet101网络下渣出钢状态识别研究

Study on the status identification of slag steel under improved ResNet101 network

针对人工目测法缺乏严格的科学依据,且受主观因素影响较大,而红外下渣检测方法需对红外图像进行去噪等预处理,易造成图像失真,并且无法提取图像颜色等基本特征,这些缺点造成最终检测准确率的降低;将空间金字塔池化与ResNet101网络相结合,提出一种基于改进ResNet101网络下渣彩色图像的出钢状态分类识别检测方法,首先将下渣检测图像结合专家经验进行分类,将数据集分为训练数据集、交叉验证数据集与测试数据集;其次,分析ResNet101网络模型,加入空间金字塔池化层(SPP-Net),便于任意大小图片的输入;最后采用Softmax分类器进行图像识别实验,钢水下渣图像识别准确率达到99%以上,能够准确实现区分所有下渣情况的功能。

The manual visual inspection method lacks strict scientific basis and is greatly affected by subjective factors.However,the infrared slag detection method needs to carry out denoising and other pretreatment on the infrared image,which is easy to cause image distortion and unable to extract basic features such as image color.Combined with the spatial pyramid pooling and the ResNet101 network,a method for the classification and detection of the tapping state based on the improved ResNet101 network bottom slag color image was proposed.Firstly,the slag tapping detection image was classified by combining with the expert experience,and the data set was divided into training data set,cross validation data set and test data set.Secondly,the network model of ResNet101 was analyzed and SPP-Net was added to facilitate the input of images of any size.Finally,softmax classifier was used for image recognition experiment,and the image recognition accuracy of molten steel slag reached more than 99%,which can accurately realize the function of distinguishing all slag conditions.