Fully Convolutional Network Variations and Method on Small Dataset

Abstract

Fully Convolutional Network (FCN) labels each pixel in an image with its category by up-sampling convolutional layer to the exact shape of input image. This paper presents a detailed evaluation on Fully Convolutional Network variations and method on small dataset. The paper mainly discusses three FCN models based on VGG16, containing FCN-32s, FCN-16s and FCN-8s, which are different in their up-sample multiple and process of fusing skipped layers. FCN based on ResNet and vanilla Convolutional Neural Network (CNN) are discussed as well for comparative experiment. Because of the small dataset, FCN method is quite different from the general, therefore arguments containing kernel size and up-sample method are tuned to increase accuracy for each kind of model. Arguments with highest accuracy are picked for comparative experiment among different kinds of model, which are FCN based on VGG16, ResNet and vanilla CNN. Mean Intersection over Union (mIoU) metric is computed as well to contrast segmentation performance among models and among classes. Loss, accuracy and mIoU after 300 epochs of training are compared. optimize processes of models are recorded to evaluate converge trend. Among all models implemented in our experiment, FCN-8s stands out, reaching the accuracy of 86.79% after 300 epochs, only by training a small dataset including 367 train images and 101 test images.