Deep learning is used in all areas of the image processing like object detection/localization, synthetic image generation, segmentation, tracking, and others. It is frequently used especially in medical image segmentation field since it provides rapid response during the treatment process. The fact that natural images contain different types of noise, patterns, and structures and the lack of distinctive quantitative information still makes the segmentation problem very challenging. The classical networks having high parameters have a long training time. The need of less training time for high parameter networks and high segmentation accuracy has led us to develop a new network. In this study, a state-of-the-art autoencoder network (MSRSegNet) is proposed to perform segmentation. Unlike conventional autoencoder approaches, it consists of encoder, fusion and decoder blocks. In encoder and decoder blocks, Multi-scale Residual Blocks are used to share information between blocks and to detect features on different scales. In fusion block, Atrous Spatial Pyramid Pooling (ASPP) module is used to preserve multi-scale contextual information. Information sharing between blocks has increased the ability of the proposed method to capture global features. The performance parameters of mean intersection over unit (mIOU) and pixel accuracy (PA) is used to compare the results. As a result, it was observed that the proposed segmentation network has high accuracy (69% mIoU) and fast segmentation performance (0.061sec. for an image with 256x256)
Discussion(0)
No comments yet. Be the first to comment.