Research And Application Of Coffee Bean Roasting Image Classification Based On Deep Learning

Coffee roasting is a critical step in coffee processing,and the quality of a cup of coffee is determined by approximately 30% of the coffee bean roasting process.Currently,smart coffee roasters can complete the roasting process semi-automatically.However,the degree of roasting still needs to be judged through an "observation window" and cannot be judged in realtime on intelligent terminals,which limits the development of intelligent coffee roasting machines.To achieve real-time monitoring of coffee bean roasting degree,this study conducted a series of research using deep learning-based image classification methods.Deep learning-based coffee bean roasting image classification algorithms face key challenges in practical applications such as sparse data sets and limited hardware resources.This study completed the following work in coffee bean roasting image data,lightweight network design,and system design.Firstly,to address the problem of sparse coffee bean roasting image data,this study proposed a coffee bean roasting image dataset for training and testing deep learning-based coffee bean roasting image classification algorithms.Considering background diversity,roasting category diversity,and viewpoint diversity,nearly 5,500 coffee bean roasting images were captured using a camera.Finally,high-redundancy and low-quality images were filtered and sorted,resulting in a dataset of 2,192 coffee bean images of various roasting states.Secondly,to address the problem of limited hardware resources in practical environments,this study proposed a lightweight network,RepFd Net,based on local attention enhancement,for coffee bean image classification under roasting conditions.In the specific structure design of Rep-Fd Net,a new frequency division module was adopted to enhance the network’s attention to local features and improve its ability to distinguish high and low-frequency information.In addition,a three-way feature fusion structure was used in the backbone network to extract fine-grained information from multiple different scales of perception field.In the inference stage,a re-parameterization method was used to fuse the threebranch structure into a single-branch structure,which ensured network accuracy while accelerating network inference speed and reducing memory usage.The Rep-Fd Net proposed in this study achieved a classification accuracy of 98.2%,meeting the classification requirements.Furthermore,its calculation amount,parameter amount,and memory usage were only25.80 MFLOPs,1.02 M,and 2.75 MB,respectively,which effectively solved the problem of limited computing resources.In terms of inference speed,it reached 124.99N/s,meeting the requirements of real-time classification in industrial applications.Finally,to achieve intelligent coffee roasting machines,this study developed a coffee bean roasting image classification system for real-time monitoring of coffee bean roasting degree.The system was developed using Pyqt5 as the development framework and included three functional modules:data storage,real-time video monitoring,and image classification.Through functional testing,the system demonstrated reliable data storage capabilities and performed real-time classification of coffee bean roasting state while presenting monitoring videos.This study provides a feasible solution for monitoring coffee bean roasting state,providing ideas for applying artificial intelligence technology to coffee roasting machines.Furthermore,this research provides technical support for promoting the intelligence of coffee roasting machines.