Image Segmentation Essay

828 Words4 Pages
Compare to the first group of neural networks, the image segmentation applications, especially for Biomedical dataset, usually spend more time on inference phase than classification applications. For example, Unet [20] will spend 3 to 4 seconds on a single image of ISBI cell tracking dataset on NVIDIA Tesla P100 GPU while AlexNet [11] can test hundreds of ImageNet [6] images within 0.1 second on the same machine. There are basically two reasons caused this phenomena. The first is the image output from electron microscopy contains about one million pixels, which is much larger than the images for the classification applications. The second is, in order to get high accuracy for biomedical data (since they are more sensitive to the…show more content…
As a result, there will be a lot of redundant computation as the pixels in nearby chunks overlap. FCN [15], Unet [20] and SegNet [1] applied similar frameworks to solve these image segmentation problems. FCN(Fully Convolution Networks) used traditional classification neural networks initially to track the features with deep, coarse, semantic data from the image, and then combined it with shallow, fine, appearance information in the last several layers of the network. The first part of the FCN determines the content in the image, while the second part combines with the location information of the objects to get the segment the input image. Unet refined the FCN architecture further. The input image in Unet traverses through several convolution layers followed by a pooling layer to get a low resolution picture. This low resolution picture is then passed to the top by the deconvolution layers thereby resulting in a u-shaped architecture. As the low resolution picture from the bottom layers traverse the upper deconvolution layers, it is concatenated with the images from the corresponding max-pooling layers. Such concatenation of original images when de-convoluted with the lowresolution images result in a high precision feature. The Unet is mainly designed for the biomedical image segmentation tasks. SegNet also employs a u-shaped architecture. The difference of SegNet

More about Image Segmentation Essay

Get Access