Efficient Convolutional Neural Networks for Pixelwise Classification on Heterogeneous Hardware Systems

摘要: This work presents and analyzes three convolutional neural network (CNN) models for efficient pixelwise classification of images. When using networks to classify single pixels in patches a whole image, lot redundant computations are carried out when sliding window networks. set new architectures solve this issue by either removing or fully that inherently predict many at once. The implementations the accessible through utility on top Caffe library. The provides support wide range image input output formats, pre-processing parameters methods equalize label histogram during training. library has been extended layers backend availability wider hardware such as CPUs GPUs OpenCL. On AMD GPUs, speedups $54\times$ (SK-Net), $437\times$ (U-Net) $320\times$ (USK-Net) have observed, taking SK equivalent SW (sliding window) baseline. throughput is up one megapixel per second. analyzed distinctive characteristics apply training processing, not every data suitable architecture. quality predictions assessed two tissue sets, which ISBI 2012 challenge set. Two different loss functions, Malis Softmax loss, were used pipeline, consisting models, interface modified library, available Open Source software under working title Project Greentea.