Improve The Fully Convolutional Network Accuracy by Levelset and The Deep Prior Method

Abstract

Deep learning techniques allow us to achieve image segmentation with excellent accuracy and speed. However, challenges in several image classification areas, including medical imaging and materials science, are usually complicated as these complex models may have difficulty learning significant image features that would allow extension to newer datasets. In this study, an enhancing technique for object detection is proposed based on deep conventional neural networks by combining levelset and standard shape mask. First, a standard shape mask is created through the "probability" shape using the global transformation technique, then the image, the mask, and the probability map are used as the levelset input to apply the image segmentation. The test results show that when using the proposed method with DCNN, it can achieve a close segmentation area and extract features with higher detail than traditional segmentation. The proposed model achieved 94.43% in precision and 95.91% in recall %, so it got 95.16% in F1-score. When comparing the proposed model with the same CNN model without Levelset, the result shows that the proposed model achieved accuracy of 0.951, which is higher than CNN model without Levelset that achieved 0.902.