Research On Tobacco Box Identification And Box Core Temperature Detection Technology

With the gradual widening of the range of incoming tobacco from each leaf regrilling unit,cross-production area and modular processing have become the norm,making the number of offline cartons of each regrilling unit increase dramatically,with a wide variety of carton identification and increasingly complex identification information.Therefore,the traditional manual quality inspection method cannot adapt to the technical requirements of seasonal mass storage,cross-production area and multispecification box identification and box core temperature detection for leaf re-baking production.In order to effectively solve this technical problem and promote the effective application of deep learning technology and industrial production automation technology in the field of leaf regrilling,this paper will start from deep learning complex scene text localization technology,recognition technology and industrial production automation technology,and carry out research on automatic identification and detection technology of tobacco box identification and box core temperature,and the main research contents are as follows:Firstly,a rotational decoupled directed text localization method based on YOLOv5 is proposed for the text localization of tobacco box identification in the process of leaf re-baking production.The method adds a parallel up-sampling and down-sampling channel to the multi-scale feature fusion by borrowing the idea of DenseNet dense connection and inheriting the YOLOv5 feature fusion network,converting the feature fusion from the original sequential stacking method to a parallel two-branch structure to enhance the image feature fusion capability.In addition,in order to solve the problem of directional text localization accuracy,a rotation decoupled bounding box detection module is designed to improve the text localization precision.Secondly,a text recognition method based on convolutional recurrent neural network(CRNN)is proposed for the recognition of tobacco box logos during the leaf re-baking production process.The method uses DenseNet,a densely connected residual network,and SE-Net,a weighted attention network for image features,to construct the SEDenseNet feature extraction network to enhance the text feature extraction capability.In the model recurrent layer,a bi-directional gated recurrent unit(Bi-GRU)with a more concise structure and fewer parameters is used to learn and model the hidden states extracted by SE-Dense Ne and the connections between spatial features to predict the preliminary sequence results,thus effectively reducing the model parameters and improving the model computational performance.In the model recurrent layer,each feature sequence predicted from the Bi-GRU output is converted into labeled sequences by CTC network for text recognition.Thirdly,In order to quickly and accurately detect the temperature of tobacco box cores in the process of leaf re-baking,a set of core temperature detection device is designed,which mainly includes: a power transmission mechanism,a multi-axis temperature measurement mechanism and a temperature isolation and noise reduction measurement rod.The multi-axis temperature measurement mechanism can effectively achieve non-interference free movement within a limited space by means of three-axis linkage,while ensuring the smoothness of punching and temperature measurement integration;the noise-isolating temperature measurement rod speeds up the digital signal acquisition speed by directly contacting the sensor with the measured object,thus improving the detection response speed and temperature accuracy.Last,combined with the actual production environment of the leaf re-baking tobacco factory,we built the whole tobacco box identification and box core temperature detection system,and verified its effectiveness and practicality.