Halftoning with Multi-Agent Deep Reinforcement Learning

Abstract

Deep neural networks have recently succeeded in digital halftoning using vanilla convolutional layers with high parallelism. However, existing deep methods fail to generate halftones with a satisfying blue-noise property and require complex training schemes. In this paper, we propose a halftoning method based on multi-agent deep reinforcement learning, called HALFTONERS, which learns a shared policy to generate high-quality halftone images. Specifically, we view the decision of each binary pixel value as an action of a virtual agent, whose policy is trained by a low-variance policy gradient. Moreover, the blue-noise property is achieved by a novel anisotropy suppressing loss function. Experiments show that our halftoning method produces high-quality halftones while staying relatively fast.

Fast halftoning policy network with error feedback

“…Additionally, to render structural details, a lightweight 2D convolutional layer is introduced to extract textural features, whose parameters are simultaneously learned with the backbone 1D CNN. The proposed policy network can converge stably under the RL-based training framework 23 without extra halftone datasets or auxiliary neural networks.…”

“…In addition to classic approaches like ordered dithering, [1][2][3][4] error diffusion, [5][6][7][8][9][10][11][12] and search-based methods, [13][14][15][16][17][18] recently, deep learning-based solutions [19][20][21][22][23][24][25] are showing their abilities in rendering decent halftones with reversibility 21 or less computational complexity. 23 Specifically, convolutional neural networks (CNNs) are trained to project white Gaussian noise maps into halftone pixels conditioning on the continuous-tone image [illustrated in Fig. 1(b)], by means of self-supervised learning, 21 unsupervised learning, 20,22 and reinforcement learning (RL).…”

“…1(b)], by means of self-supervised learning, 21 unsupervised learning, 20,22 and reinforcement learning (RL). 23,25 This paradigm enables one to develop halftoning algorithms in a more automatic manner, in which the processing efficiency of CNNs is backed by modern high-performance computing platforms, e.g., graphic processing units (GPU) and neural processing units.…”

“…Section 2 provides a brief but necessary review of RL-based halftoning. 23 In Sec. 3, the proposed model is introduced in detail.…”

Diverse Effective Relationship Exploration for Cooperative Multi-Agent Reinforcement Learning

“…One of the key challenges in this field is the implementation of Due to the generalization and applicability of reinforcement learning algorithms, they are increasingly being applied to inventory management problems. Examples include Deep Q-Network (DQN) [50], QMIX [51], QTRAN [52], IPPO and MAPPO [53], and CD-PPO [54]. These RL algorithms demonstrate promising capabilities for addressing inventory management challenges and may provide performance improvement and better adaptability.…”