基于深度强化学习的木材缺陷图像重构及质量评价模型研究

Research on wood defect image reconstruction and quality evaluation model based on deep reinforcement learning

针对典型仿生智能算法处理木材缺陷图像感知及质量决策问题时存在的多维退化因素作用下的缺陷图像失真严重、缺陷图像先验特征提取方差波动频繁、质地不均匀缺陷图像灰度分割失效、异种木材自身纹理泛化能力与学习能力失衡、最优收敛速度随缺陷维度呈迟滞变化等先天不足,提出了一种基于深度强化学习的木材缺陷图像重构及质量评价模型。引入深度学习机制,通过利用深度残差网络进行迭代训练,实现差异性木材多维缺陷图像实时高效重构,构建面向差异性木材多维缺陷精细分割与特征提取的全景自主感知模型,构建大数据量级木材缺陷特征共享资源池;引入强化学习机制,利用深度确定性策略梯度算法建立缺陷特征迭代更新、自主决策、全景可视、深度预测与木材质量评价之间的高维度决策映射,实现多维差异性木材缺陷图像重构及质量评价的横向共享集成。基于Tensorflow开源框架,在Gym Torcs环境下进行模型效能仿真验证,较好解决了典型仿生智能算法处理木材缺陷图像感知及质量决策问题时存在的若干固有缺陷,实现木材缺陷图像自主感知重构及质量评价自主决策,具有缺陷特征感知全面、抗干扰性强、自主决策性高等优势。以浙江省湖州市南湖林场辖区内某经济林木为效能评价载体,对模型进行了工程应用分析,结果表明,该模型可以较好实现木材多维缺陷感知与重构、全局最优质量评价自主智能决策,在感知自主性、重构复现性、自主决策性、模型泛化能力等方面具有明显优势。

Aiming at the problem of wood defect image perception and quality decision-making in typical bionic intelligent algorithm,the defect image distortion was serious,the variance of prior feature extraction of defect image fluctuates frequently,the gray level segmentation of defect image with uneven texture is invalid,the generalization ability and learning ability of different wood texture are unbalanced,and the optimal convergence speed is delayed with the defect dimension,a model of wood defect image reconstruction and quality evaluation based on deep reinforcement learning was proposed.By introducing the deep learning mechanism and using the deep residual network for iterative training,we can realize the real-time and efficient reconstruction of the multi-dimensional defect image of different wood,build a panoramic autonomous perception model for the fine segmentation and feature extraction of multi-dimensional defect of different wood,and build a large data level shared resource pool of wood defect features;By introducing reinforcement learning mechanism and using depth deterministic strategy gradient algorithm,a high-dimensional decision mapping among iterative updating of defect features,independent decision-making,panoramic visibility,depth prediction and wood quality evaluation was established,which realized the horizontal sharing integration of multi-dimensional difference wood defect image reconstruction and quality evaluation. Taking an economic forest in Nanhu forest farm area of Huzhou city,Zhejiang province as the evaluation carrier,the engineering application analysis of the model was carried out. The verification results showed that the model proposed in this paper can better realize the multi-dimensional defect perception and reconstruction of wood,the autonomous intelligent decision-making of global optimal quality evaluation,and has the obvious ability of sensing autonomy,reconstruction reproducibility,autonomous decisionmaking,model generalization,etc show superiority.