Using a Neural Network Method to Solve Image Segmentation Problems

Abstract

Image segmentation plays an important role in detecting various diseases and pathologies through medical image processing. Over the years, a number of traditional approaches such as the binary threshold method (Otsu method), watershed method, and K-means clustering method have been developed to perform image segmentation, using domain-specific knowledge to efficiently solve segmentation problems in specific applications. But unfortunately, these methods were not effective. In fact, the essence of image segmentation is the classification of containing each pixel with similar attributes. And the goal of classifying a pixel is to treat it coarse, insensitive to noise, fine, and a medium star filter. Something that traditional methods cannot do. Therefore, there is a need to propose a new method that allows eliminating these disadvantages. We propose a method based on autoencoder (a special architecture of artificial neural networks that allows the application of learning without a teacher using the method of back propagation of the error), which allows compressing any large image into a small one, (with the same properties as the input). This method also saves resources. In this case, we use an image processing method called "MaxPooling" (U-Net). In the work, we created our own metric to track the result of the network training. In the practical part, we tried segmenting into classes the images from the construction site, as well as pixel by pixel the locations of the aircraft in the images.

Keywords

• segmentation
• pixel
• autoencoder
• maxpooling
• u-net
• metrics
• object detection
• artificial neural networks
• cnns
• full-width network