Computer Vision with Deep Convolutional Neural Network Approach for Cold-Flow Casting Defect Detection
Quality of the engine parts is very important, because will be directly affected to motorcycle performance. If the defects from the Aluminum High-Pressure Die Casting (HPDC) injection process that are not detected and sent to the next process will cause production loss and will harm the customer. The research presents the implementation of object detection technology based on Deep Convolutional Neural Network (DCNN) to detect cold-flow defect. Computer vision with DCNN algorithm will be used to improve the result of visuals human inspections who have been detecting cold-flow defects that will in a result more stable and objective in carrying out quality assessments. This research will analyze the performance of the DCNN framework called YOLOv5s in Python. The analysis includes lighting conditions and the characteristics of the dataset. This research used cloud computing at Google-Colab during the training process of the DCNN, the computer specifications with graphic processing unit known as GPU (Tesla T4, 1.5 GB, 40 processor assigned by Google-Colab) are needed for faster training process. The Roboflow was very helpful tools in the dataset preparation phase of the development of this system. In conclusion, the developed system has proven to be very successful in assisting the HPDC part visual inspection, with mAP value 0.33, box loss value 0.06 and the average detection speed per object defect is 0.3 seconds.