基于深度强化学习的木材缺陷图像识别及分割模型研究

Research on wood defect image recognition and segmentation model based on deep reinforcement learning

针对典型仿生智能算法处理木材缺陷图像精确识别及最优分割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识别及分割模型。引入深度学习机制,利用深度卷积神经网络进行迭代训练,实现差异性木材多维缺陷图像特征实时高效提取,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,利用双重Q网络机制建立缺陷特征迭代更新、自主决策、全景可视、深度预测与缺陷图像识别之间的高维度决策映射,实现多维差异性木材缺陷图像精确识别及最优分割的横向共享集成。基于PyTorch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精确识别及最优分割问题时存在的若干固有缺陷,实现木材缺陷图像精确识别及最优分割,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,验证结果表明所提模型可以较好实现木材缺陷图像精确识别及最优分割,在感知自主性、最优收敛速度、分割全局最优性、缺陷图像保真度等方面具有明显优势。

For the typical bionic intelligent algorithm to deal with the accurate identification and optimal segmentation of wood defect images,the defect image is seriously distorted under the action of multi-dimensional degradation factors,the priori feature extraction of defect images fluctuates frequently,the uneven texture has uneven grayscale segmentation,the heterogeneous wood′s inherent texture generalization ability and learning ability are unbalanced,and the optimal convergence speed varies lagging along with the defect dimensions.The wood defect image recognition and segmentation model based on deep reinforcement learning is proposed.Introduce deep learning mechanism and use deep convolutional neural network for iterative training to achieve real-time and efficient extraction of differential wood multi-dimensional defect image features,build a panoramic autonomous perception model for fine segmentation and feature extraction of differential wood multi-dimensional defects,and construct large data magnitudes wood defect feature shared resource pool,introduction of reinforcement learning mechanism,using dual Q network mechanism to establish high-dimensional decision mapping between defect feature iterative update,autonomous decision-making,panoramic visualization,depth prediction and defect image recognition to achieve multi-dimensional differential wood defects Horizontal image sharing for precise image recognition and optimal segmentation.Based on the PyTorch open source framework,model efficiency simulation and verification under the Gym Torcs environment has solved some inherent defects in the typical bionic intelligent algorithm when processing wood defect image accurate recognition and optimal segmentation,and realized wood defect image accurate recognition and excellent segmentation has the advantages of comprehensive defect feature perception,strong anti-interference and high self-determination.Using an economic forest tree in the area of Nanhu forest farm in Huzhou city,Zhejiang province as the carrier of efficiency evaluation,the model was analyzed for engineering applications.The verification results show that the model proposed can better achieve the accurate identification and optimal segmentation of wood defect images,and it is self-conscious.The optimal convergence speed,the global optimality of segmentation,and the fidelity of defective images have obvious advantages.