YOLOX-Based Design Of An Embedded Inkjet Inspection System

China has announced a number of regulations and actions in recent years to direct businesses toward creating goods with intelligence and technical innovation.The Intelligent Manufacturing Development Plan also outlines the requirements for intelligent production for businesses,encourages joint venture development and labor sharing among businesses in each value chain component ring,and gradually establishes the framework for an intelligent production system integrator.Product packaging information processing technology has achieved full high-speed automation,but the related quality inspection link is still carried out using the human quality inspection method.Currently,the quick automatic manufacturing does not match the manual quality checking approach.The problem of the speed gap between automatic production and manual quality inspection will be resolved by the application of automatic machine vision identification technology in product information quality inspection,which is the result of thorough research and application of deep learning in the field of machine vision.The hardware cost of the currently available inspection systems on the market is relatively high because the industrial computers with x86 architecture that are frequently used in industry serve as the system’s main processing unit.The majority of the remaining small and medium-sized enterprises are unable to purchase the equipment,with the exception of a few well-known large manufacturers.The following are the primary research topics covered in this paper:(1)Use YOLOX’s network as the benchmark model to enhance and optimize,and in this paper’s inkjet information placement algorithm,utilize the improved algorithm.To reduce network parameters and calculations,the main optimization task entails cutting the branch of the output network structure of the network head.The CA attention mechanism is also introduced to help the network become more focused on the characteristics of the black inkjet dots and the YOLOX network become more focused on the target object that needs to be identified for this task.Because CIo U takes into account the distance and similarity between the ground true boxes and the predicted boxes,and more precisely locates the information area of the inkjet code,it performs better than the position regression loss function of the original network and has a faster convergence rate during network training.The network channel was shortened and the streamlined YOLOX-s network became lightweight once more for the following deployment of embedded devices.(2)Using the CRNN character recognition algorithm as the benchmark model,the original network’s structure is trimmed and changed in light of the inkjet code identification task’s complicated backdrop,relatively constant string length,and limited character styles.The enhanced CRNN network on the embedded platform has a inference time of 5.78 ms and a sequence accuracy of 98%.(3)Prior to deploying the model to an embedded platform,model optimization is required when YOLOX-s and CRNN are combined.The aforementioned approach is implemented on the NVIDIA Jetson TX2 using the Tengine AI framework to provide realtime inkjet data examination on embedded platforms.QT Creator software is used to create a straightforward user interface for the system in order to make it low threshold and simple to use.The system may be operated in accordance with the interface’s instructions,and the interface can show and track product quality inspection statistics in real time.(4)The NVIDIA Jetson TX2 is chosen as the system’s core module based on the enterprise’s evaluation of production equipment cost control in advance and product scalability,and a set of real-time inkjet information detection and identification systems is constructed.The system is then debugged and put through a performance test on the actual production line to make sure it can accurately and quickly carry out the aforementioned responsibilities of information quality inspection.The actual measurement reveals that this system’s recognition rate can reach up to 98.1% when the product’s pipeline motion is less than or equal to 25 / s.The testing findings show that the system has essentially complied with quality control standards in real production.