基于知识蒸馏的轻量化残差网络在塑料颗粒高速检测中的运用

Application of Lightweight Residual Network Based on Knowledge Distillation in High Speed Detection of Plastic Particles

为了实现塑料颗粒生产过程中的光学外观高速检测,构建了一种基于多教师模型融合的知识蒸馏轻量化残差网络,用多个教师融合模型指导学生网络训练。以结构较为复杂的Vgg-16、Vgg-19、Resnet-18网络为教师网络,以自主搭建的具有多层串联小卷积、残差块特点的轻量化残差网络为学生网络,通过知识蒸馏训练塑料颗粒缺陷检测网络。为了增加模型的泛化能力及鲁棒性,在激活函数的选择上采用LeakyRelu解决了梯度消失的问题,在梯度下降算法的优化方式上采用余弦退火跳过了局部最优点。试验结果表明,轻量化残差网络经过知识蒸馏后与未经过相对比,精确度和召回率分别提高了1.8%和2.2%,精确度从88.8%提升到90.6%,召回率从96.7%提升到98.9%。学生模型经过知识蒸馏后基本达到了教师模型的性能水平,模型的参数量、浮点运算次数、单次检测耗时分别降低了59.2%、59.124%、50%。提出的知识蒸馏模型通过自主搭建的轻量化残差网络在保证检测精度的基础上,实现了塑料颗粒的高速检测,为深度学习技术在塑料颗粒在线高速全检的产业化应用提供了一种方法。

Aiming at the high-speed detection of optical surface in the production process of plastic particles,a lightweight residual network of knowledge distillation based on multi teacher fusion model was established,and multi teacher fusion model was used to guide students'network training.The plastic particle defect detection network was trained through knowledge distillation model,which was taking Vgg-16,Vgg-19 and Resnet-18 networks with complex structure as the teacher network and establishing lightweight residual network with the feature of multi-layer small convolution connection and residual block as the student network.In order to increase the generalization and robustness of the model,LeakyRelu was used to solve the problem of gradient disappearance as activation function,and cosine annealing was used to skip the local optimum in the optimization method of gradient descent algorithm.The results show that after knowledge distillation,the accuracy and recall of lightweight residual network are improved by 1.8%and 2.2%respectively,the accuracy is improved from 88.8%to 90.6%,and the recall is improved from 96.7%to 98.9%.After knowledge distillation,the student model basically reached the performance level of the teacher model.The parameters,floating-point operation times and single detection time of the model are reduced by 59.2%,59.124%and 50%respectively.The proposed knowledge distillation model realizes the high-speed detection of plastic particles on the basis of ensuring the detection accuracy through the self built lightweight residual network,which provides a method for the industrialized application of deep learning technology in the on-line high-speed full detection of plastic particles.