Abstract

To solve the problem of fabric defect detection under complex illumination conditions, the Recurrent Attention Model (RAM) which is insensitive to illumination and noise differences has been introduced. However, the policy gradient algorithm in the RAM has some problems, such as the difficulty of convergence and the inefficiency of the algorithm due to the shortcomings of round updating. In this paper, the Deep Deterministic Policy Gradient- Recurrent Attention Model (DDPG-RAM) algorithm is proposed to solve the problems of policy gradient algorithm. Although the decoupling of the reinforcement learning task and classification task will lead to the inconsistency of the data, the gradient variance will be smaller, and the convergence speed and stability will be accelerated. Experiment results show that fabric defects can be detected by the proposed DDPG-RAM algorithm under complex illumination conditions. Compared with RAM and the Convolutional Neural Network (CNN), the accuracy of the decoupled algorithm is 95.24%, and the convergence speed is 50% faster than that of the RAM.

摘要

为了解决复杂光照条件下的织物疵点检测问题，引入了对光照和噪声差异不敏感的循环注意模型（RAM）。但是RAM中的策略梯度算法由于轮更新的缺点，存在收敛困难、算法效率低下等问题。本文提出了深度确定性策略梯度-循环注意力模型（DDPG-RAM）算法来解决策略梯度算法的问题。强化学习任务和分类任务的解耦虽然会导致数据不一致，但梯度方差会更小，收敛速度和稳定性会加快。实验结果表明，本文提出的DDPG-RAM算法可以在复杂光照条件下检测织物疵点。与RAM和卷积神经网络（CNN）相比，解耦算法的准确率为95.24%，收敛速度比RAM快50%。