Design And Implementation Of Embedded Inkjet Detection System Based On Deep Learning

The widespread adoption of industrial production automation technology has considerably increased production efficiency.The conventional quality inspection method has been unable to fulfill the current production demand as the date of manufacturing,shelf life of products,and other information printing technology have matured.Using computer vision technology to solve the inspection task of inkjet coding will become a hot topic in the research of many manufacturers as deep learning technology becomes more widely employed in the field of industrial computer vision.Existing computer vision systems for inkjet character inspection are all based on industrial computers or computers with GPU devices,which are not conducive to the detection system's commercialization in terms of volume,cost,power consumption,or adaptability.Based on this possibility,this study investigates a deep learning-based character recognition approach for point inkjet coding and chooses the NVIDIA TX2 as a low-cost,small-volume integrated platform to test the improved algorithm.The following are the research's main points:(1)Investigate the inkjet character location method using the YOLOv5 s network as a baseline model.The ECA attention mechanism is introduced to increase the network's feature extraction capabilities and to alleviate the algorithm's missing detection problem.To increase the accuracy of location,the aspect ratio factor parameter is incorporated into the network's rectangular frame regression loss function.Finally,to execute sparse training and structured pruning of the YOLOv5 s network model,a structured pruning method is used,which decreases the model's complexity and reasoning time.The improved network model's positioning accuracy is 99.9% and 93%,respectively.(2)The network structure of the inkjet character is simplified using CRNN's inkjet character recognition algorithm,and the entire connection layer is used to replace the complex LSTM structure to increase the character recognition model's reasoning speed.The CRNN-Lite network's character recognition rate has improved to 97.15%,with a PC inference time of only 3.1 milliseconds.(3)The YOLOv5+CRNN network model was converted to a different format and its parameters were quantified.The network was deployed to the NVIDIA TX2 embedded platform using the Tengine AI reasoning framework to complete the embedded reasoning phase of the inkjet coding character detection algorithm.To complete the software design of the inkjet detection system,C++ and QT5 were used to construct a basic human-computer interaction interface to make it easier to use.A deep learning technique was used to display the data from pipeline inkjet detection as well as the system's control operation in real time.(4)A set of image acquisition and detection systems was designed for the character printing environment in the industrial industry using the NVIDIA TX2 embedded platform.To ensure the system's viability,the enhanced algorithm was tested in a field printing scenario.When printing at a rate of less than 25 frames per second,the recognition rate can exceed 97%.The results reveal that the system is capable of meeting the actual demand for jet code checking on the production line.