Image Classification of Zinc Dross Based on Improved MobileNetV2

Zinc dross image classification is one of the key procedures to realize the automation of dross removal. Given the low efficiency of traditional deep convolutional neural networks for image classification, the improved MobileNetV2 image classification lightweight network is proposed for achieving real-time and high-precision classification of zinc dross images in this paper. The improvement includes three aspects. Firstly, the batch normalization method in MobileNetV2 is replaced by the group normalization method. Then, the Relu6 activation function in the network is changed to the Leaky Relu. Thirdly, the expansion coefficient used to control the dimension elevation is decreased, which makes the output channels of the pointwise convolution layer less. Comparison experiments were performed on the enhanced zinc dross image data set. Finally, the experimental results show that the parameter amount of the improved MobileNetV2 is just 1.33 million, which is 41.67% less than MobileNetV2. The improved MobileNetV2 has a 99.63% accuracy in the classification of zinc dross images, which is 1.46% higher than MobileNetV2. And the test time of the improved MobileNetV2 is 8.49 ms, which is 7.57 ms less than that of MobileNetV2.