Mold ID Recognition Method Based On Differential Neural Network Architecture Search

With the rapid development of deep learning,the object detection method based on deep learning has gradually replaced the traditional object detection method,and has made great progress.Many object recognize methods based on deep learning have achieved excellent performance on various public data sets.However,when these methods are used directly for specific tasks(E.g.mold code recognition task),problems such as insufficient task data sets and insufficient accuracy of the detection model for hyperparameter tuning for specific tasks will happen In addition,the structure of the object detection network based on deep learning is complicated,and manual optimization of the network hyperparameters requires a lot of expert experience as well as a large amount of computing resources for network training and verification.In order to solve the problem of insufficient task data set,this paper proposes a mold ID recognition method based on Mask RCNN.Optimize the anchor frame design for the mold code recognition data set and use pre-training to improve the performance of the model.Aiming at the complex problem of network hyperparameter tuning,this paper proposes a Mask RCNN hyperparameter optimization method based on Differentiable Neural Network Architecture Search(DARTs).First,we define the network cell structure.Then we use DARTs to train the network cell structure.Finally,the trained network cells are used to reconstruct the classification and regression branches in the Mask RCNN network to optimize the Mask RCNN object detection network for specific tasks.The method of this paper was tested on the mold code recognition data set we constructed.The average accuracy of proposed method is 1.96% higher than that of traditional Mask RCNN object detection method with Io U=0.5.The average accuracy of the proposed method is 4.83% higher than that of traditional Mask RCNN object detection method with Io U=0.75.The overall character accuracy of proposed method is 7.58% higher than that of traditional Mask RCNN object detection method with Io U = 0.5 and Io U = 0.75.In addition,this method can converge under 2GPU Days,and its computational complexity is not high.The main contributions of this paper are as follows:(1)We make a mold code recognition data set containing 7065 images and analyze the data set.The data set can be downloaded and used at https://www.kaggle.com/hewaele/dataset-of-moldid.(2)We propose a mold code recognition method based on Mask RCNN.For the identification task of mold coding,the anchor frame in Mask RCNN is optimized,and the pretraining strategy is used to improve the performance of the model.(3)We propose a method to optimize Mask RCNN based on DARTs.By using DARTs to reconstruct the classification and regression branches in Mask RCNN,we optimize the model.The effectiveness of the proposed method is verified through multiple experiments with mold code recognition data set.