Few-shot and Fast Texture Segmentation Based on Non-padding Convolution

To solve the segmentation problem of textures, the accuracy of classical algorithms of machine vision is not satisfactory because easily confused textures are often difficult to distinguish accurately. However, state-of-art semantic segmentation algorithms based on deep learning require expensive equipment, long time and large amounts of data for training, and large human resources for data annotation. This paper is committed to building a simplified convolution model based on non-padding convolution, which requires a very small number of images to train, and it can realize fast texture segmentation on the CPU device. On the dataset used in this paper, FB-IoU can reach 92.8%. On the 2.3GHz CPU, the average processing time of
$1920\times 1080$
images is 428ms. The accuracy has obvious advantages over classical segmentation algorithms or few-shot segmentation algorithms based on meta learning, and the processing speed has obvious advantages over lightweight semantic segmentation models based on deep learning.