Reducing the Number of Trainable Parameters does not Affect the Accuracy of Res-Net18 on Cervical Cancer Images
K. Sreenivasa Rao, Priyadarshini Chatterjee · 2024
Res-Net18 is one amongst the most important tool for medical image processing. It is used both for image segmentation and classification. There are CNN that are used to train very deep network using the ‘technique of batch normalization. However, these CNNs cannot tackle the problem of degradation. Researchers have found out an improved version of CNN called Res-Net with shortcut connection. The residual unit of this u-net helps in improving the flow of information. It makes the layers to learn according to the residual function and according to the input given in the layers. The short circuit connection makes the output and the input dimensions similar. There are various evolved versions of res u-net. This paper has considered res-net18 whose number of trainable parameters is eleven million with a standard accuracy of 94% with respect to the three datasets of cervical cancer. We have tried to reduce the number of trainable parameters of Res-Net18 by using an additional max-pooling layer in the architecture of the Res U-Net. We have successfully reduced the trainable parameters to eight million, keeping the accuracy to 94% on a cervical cancer data set.