Effective Approaches for Improving the Efficiency of Deep Convolutional Neural Networks for Image Classification

Identificador:  TDX:4167

Handle: https://hdl.handle.net/20.500.11797/TDX4167

Autors:  Schwarz Schuler, Joao Paulo

Resum:
This thesis presents two methods for reducing the number of parameters and floating-point computations in existing DCNN architectures used with image classification. The first method is a modification of the first layers of a DCNN that splits the channels of an image encoded with CIE Lab color space in two separate paths, one for the achromatic channel and another for the remaining chromatic channels. We modified an Inception V3 architecture to include one branch specific for achromatic data (L channel) and another branch specific for chromatic data (AB channels). This modification takes advantage of the decoupling of chromatic and achromatic information. Besides, splitting branches reduces the number of trainable parameters and computation load by up to 50% of the original figures in the modified layers. We achieved a state-of-the-art classification accuracy of 99.48% on the Plant Village dataset. This two-branch method improves image classification reliability when the input images contain noise. In DCNNs, the parameter count in pointwise convolutions quickly grows due to the multiplication of the filters and input channels from the preceding layer. To handle this growth, the second optimization method makes pointwise convolutions parameter-efficient via parallel branching. Each branch contains a group of filters and processes a fraction of the input channels. To avoid degrading the learning capability of DCNNs, we propose interleaving the filters' output from separate branches at intermediate layers of successive pointwise convolutions. We tested our optimization on an EfficientNet-B0 as a baseline architecture and made classification tests on the CIFAR-10, Colorectal Cancer Histology, and Malaria datasets. For each dataset, our optimization saves 76%, 89%, and 91% of the number of trainable parameters of EfficientNet-B0, while keeping its test classification accuracy.