Research On Fabric Defect Detection Algorithm Under Complex Illumination Environment Based On Reinforcement Learning And Recurrent Model Of Attention

In the textile industry,the problem of quality control has a huge impact on the profit of textile enterprises.The intensity and brightness of the light source in the vision system will have different degrees of attenuation with the time used.The traditional fabric defect detection algorithm is aimed at a certain fabric under normal lighting conditions.The accuracy of the algorithm will drop dramatically when the fabric type or illumination intensity changes.Although the features extracted by the convolutional neural network possess the versatility and extreme robustness,experimental results show that the accuracy of the model used is not high enough for fabric defect detection under complex illumination conditions.However,the recurrent attention model is not sensitive to illumination changes and noise,and is tightly coupled to reinforcement learning.Therefore,this thesis proposes a fabric defect recognition algorithm based on recurrent attention model.For the problem of complex illumination,this thesis proposes the multi-scale Retinex image enhancement algorithm to preprocess the fabric image.It is found that the gray scale transformation method will cause the overall brightness of the processed image to be too bright,and the histogram equalization algorithm will increase the noise and cause the lack of details.The multi-scale Retinex algorithm can limit the influence of illumination variation and noise while maintaining the local details of the image.Finally,experiments have proved that the multi-scale Retinex algorithm has good robustness and versatility.There are some problems about the deep deterministic policy gradient algorithm,such as the lack of scientific theoretical method guidance and the rough evaluation method which causes the small difference between the value function of the optimal action and the non-optimal action.This thesis proposes the DDPG-OSPC algorithm,which uses the temporal difference method based optimized sampling algorithm and multi-dimensional evaluation method to solve the above problems.Experiments based on the Reinforcement learning environment proved that this algorithm can meet the requirements of 25 rounds ahead of the deep deterministic policy gradient algorithm,and the peak value of the accumulated rewards and the stable value are higher.The policy gradient algorithm used in the recurrent attention model is difficult to converge,and the shortcomings of the episode update result in the inefficiency of the algorithm.This thesis proposes the DDPG-RAM algorithm,which uses the deep deterministic policy gradient algorithm to solve above problems.It is also found that whether the reinforcement learning task and the classification task are coupled has an important impact on the experimental results.The coupling will lead to a large variance of the gradient,and the decoupling will result in the data being inconsistent.Experiments proved that the DDPG-RAM algorithm proposed in this paper can do the defect detection of fabric under complex illumination conditions.Compared with the recurrent attention model and the convolutional neural network,the accuracy of the algorithm after decoupling can reach 95.24%,and the convergence speed is 50% faster than the recurrent attention model.This illustrate that the data inconsistency has less impact than the gradient variance,and the gradient variance is small,the fluctuation of the neural network will be small,the convergence speed will be faster and the stability will be stronger.