MobileNet investigation: its application and reproducing edge detectors using depth-wise separable convolution

MobileNet has a wide range of applications, and its depth-wise separable convolution deduced the computational cost dramatically compared to traditional convolution methods. In this study, experiments are carried out to examine depthwise separable convolution’s ability in terms of reproducing edge detector kernels. MobileNet’s depth-wise separable convolution achieves great performance through reproducing kernels based on our LpP metric. The visualization results show that random initialization can reduce the calculation amount of the model and have higher efficiency. This study also compares three models, including InceptionV3, ResNet50, and MobileNetV2, on face mask detection tasks. The experiment result indicates that the MobileNetV2 model achieved the highest accuracy of 99.08%. A ground truth analysis on a test image containing three different face mask situations is performed to further discuss the model performance of MobileNetV2.