Computer Vision News - June 2019

Figure 2: An illustration of the proposed MRN for two resolutions. Dark blue boxes: stacks of two convolution layers with ReLU activations. Red boxes: max-pooling layers. Light blue boxes: convolution layers with identity activations. Green boxes: transposed convolution layers with ReLU activations. Figure 3: An illustration of patches with the same central coordinates. Increased mpp values corresponds to zooming out action to enlarge the field of view. Yellow squares represent the effective tissue area at different magnifications. In a more general scope, the question regarding the balance between global and local information in visual perception is extensively studied by psychologists and neuroscientists since the rising of the gestalt psychology. In line with this, multiple deep learning researchers using networks which mimick the visual system have addressed this issue and designed networks enabling the processing of images using multi-scale data . The U-Net based MRN architecture allows the processing of local information using its global context while remaining the number of weights relatively compact. The small number of weights, which increases linearly with number of resolutions fed to the network, enables the use of a small dataset. In one of our histological image segmentation projects at RSIP Vision , we were inspired by this architecture and expanded the standard U-Net to process multiple resolutions . This approach was simple to implement and increased the network’s prediction accuracy . Research 12 Research Computer Vision News