| Surface roughness criteria for parts are becoming increasingly strict due to the advancement of modern manufacturing technologies.Therefore,surface roughness measuring technology has become essential to the growth of the manufacturing industry.In recent years,machine vision detection has become one of the popular non-contact measurement methods due to its high detection efficiency,low cost,and support for in-line detection.As a result,this method has also been gradually introduced into the field of roughness measurement.However,the current machine vision roughness detection still has some key issues that have not been effectively addressed.(1)Roughness vision detection methods based on feature index design: Although these methods require a smaller sample size,the process of designing indices artificially demands more time and labor costs.This drawback makes it difficult to quickly build the model.Moreover,this approach has issues with complicated index construction,a demanding environment for picture acquisition,and limited universality.(2)Roughness vision detection method based on neural network: In contrast to feature index detection methods,neural network-based methods can not only use neural networks to extract features automatically but also have good prediction performance.However,their advantages depend on large amounts of data and intricate model structures.More labor,time,and economic costs are required during the sample preparation and training processes in the model,which can impact both the speed and cost of model construction.In response to the aforementioned problems,this thesis presents the following research and explorations,using milling surface roughness inspection as an example.(1)A vision detection method is proposed for milling surface roughness grade based on the broad learning system.This approach aims to improve the speed of model creation by reducing model complexity.The experimental results show that this method can automatically extract the image features associated with the roughness parameters and has an high model training speed.Under normal lighting conditions,the method also exhibits high detection accuracy,good overall classification performance and sample sorting ability.(2)A vision detection method is proposed for milling surface roughness grade based on the few-shot learning network MAML++.This approach aims to accelerate model building and reduce model building costs by addressing the few-shot problem.The experimental results demonstrate that this method can identify classes of milling surface roughness with a limited number of labeled samples and identify new classes without additional training.It also has the capacity to automatically extract features,high accurate detection,and strong robustness to the lighting environment.(3)A vision-based system for milling surface roughness detection has been developed using Py Charm and Py Qt.The system can complete the entire roughness detection process,including connecting and controlling the camera,capturing images,displaying images in real-time,saving images and results,and detecting roughness.In conclusion,the research work in this thesis is an extended study based on current machine vision roughness detection methods.This study provides new and improved ideas for roughness vision detection technology,which can help promote the application and development of machine vision in the field of roughness measurement. |
