Mask Defect Detection Algorithm Based on Improved EfficientNetV2

Abstract

Aiming at the problem of insufficient detection accuracy of mask defects with many types and large differences, a deep learning classification algorithm based on improved efficientnetv2 is proposed to achieve efficient detection of fourteen complex mask defects. In this paper, efficientnetv2 with strong feature extraction ability is used as the backbone network, combined with the improved compression and incentive attention mechanism, h-swish activation function and label smoothing technology, to enhance the attention of the model to defects, improve the detection speed of the model, reduce the impact of noise, and reduce the complexity of the model. The generated model realizes the classification and recognition of mask surface defects and structural abnormalities, with an average accuracy of 98.95% and a transmission frame rate of 40fps per second.