Novel Deep Learning Approaches: Robust Learning from Noisy Labels and Multi-Task Learning for Efficient Image Processing

Abstract

Recently, deep neural networks have shown exceptional performances in several computer vision applications. However, deep learning has two main limitations for use in the industry: high complexity and vulnerability against noisy labels, also known as erroneous labels. In this dissertation, I study on improving robustness and efficiency of deep neural networks in classification and image enhancement applications. Firstly, a novel criteria is proposed to robustly train deep neural networks with noisy labels. If the labels are dominantly corrupted by some classes (these noisy samples are called dominant noisy labeled samples), the network learns dominant noisy labeled samples rapidly via content-aware optimization and they cause memorization (reduce generalization) in the deep neural network. To mitigate memorization of noisy labels, algorithm with proposed criteria penalizes dominant noisy labeled samples intensively through inner product of class-wise penalty labels and their prediction confidences, which indicate the probability of being assigned to each class. By averaging prediction confidences for the each observed label, I obtain suitable penalty labels that have high values if the labels are largely corrupted by some classes. Additionally, the penalty label is compensated using weight to make it same for learning speed per class, and it is updated using temporal ensembling to enhance the accuracy. The proposed criteria can be easily combined with the algorithms of loss correction and hybrid categories through a simple modification to improve learning performance. Secondly, multi-task learning based deep neural network is proposed to train various image processing operators efficiently. For real-time image processing, the proposed algorithm takes a joint upsampling approach through bilateral guided upsampling. For multi-task learning, the overall network is based on an encoder–decoder architecture, which consists of encoding, processing, and decoding components, in which the encoding and decoding components are shared by all the image processing operators. In the processing component, a semantic guidance map, which contains processing information for each image processing operator, is estimated using simple linear shifts of the shared deep features. Through these components, the proposed algorithm requires an increase of only 5% in the number of parameters to add another image processing operator and achieves faster and higher performance than that of deep-learning-based joint upsampling methods in local image processing as well as global image processing.