Highly Contrast Image Correction for Dim Boundary Separation of Image Semantic Segmentation

Abstract

The efficiency and accuracy of the image semantic segmentation algorithm represent a trade-off relationship, and the loss of accuracy tends to increase as the model structure simplifies to improve efficiency. Developing more efficient and accurate algorithms requires methods to complement them. In this study, we applied the logarithmic-exponential mixture (LEM) function for gamma correction of images to improve the accuracy of image semantic segmentation. The basic model used in this work was produced by constructing a full convolution neural network based on MobileNetV2. To avoid the noise of input compression, we corrected training and validation images with gamma from 1/8 to 8 (7 different levels) before doing convolution. We evaluated models using Tensorflow deep-learning library based on Python. We compared models using LEM function to models using conventional gamma function. The prediction masks of the proposed model using the LEM function had relatively small fluctuations of accuracy upon gamma change. For images that have shadows overlapped on the object, the object was better distinguished in small gamma values. For dark images, the increase in accuracy was more effective. The results indicated that the proposed gamma correction could improve image segmentation accuracy in images with unclear edges. We believe that the presented results will guide further studies for accuracy improvement of image recognition algorithms applicable to future devices, such as autonomous vehicles and mobile robots.