The effect of Compressor-Decompressor Networks with different image sizes on Mask Detection using Convolutional Neural Networks - VGG-16

Since a pandemic such as COVID-19 entered our lives, the importance of using face masks against such epidemics was experienced once again. Although the use of vaccines increasing the immunity against the COVID-19 virus prevents it from being as effective as in the early times, the use of face masks and their detection in crowded areas may still be important because another pandemic can occur any time. In this research, a preliminary work about a CNN-based compression-decompression model has been studied. In this two-staged model, first, the input image is compressed and its size is reduced to $8\times 8$ from $32\times 32,\ 64\times 64$, or $128\times 128$. In the second stage, this $8\times 8$ image is decompressed and reconstructed to its original size of $32\times 32,\ 64\times 64$, or $128\times 128$ depending on the input image size. Since the results are promising, such a system can be used to compress the image when it is acquired and then transfer it to another machine with better specifications to decompress and use more complex models for better classification accuracies. 90.9%, 91.7%, and 92.1% accuracies were obtained with the VGG-16 model by using $32\times 32,\ 64\times 64$, and $128\times 128$ images respectively. If a balance between network size and accuracy is considered, $64\times 64$ images are the optimal solution, but if the goal is better performance, $128\times 128$ images can be used to have higher accuracy by using higher-cost machines.