Accuracy Improvement of CNN MobileNet-V1 and Residual Network 50 Layers Models using Adam setting for Car Type Classification
Muhammad Taqiyuddin, Yudi Eko Windarto, Wahyul Amien Syafei, Ike Pertiwi Windasari, Agung Budi Prasetijo · 2022
The industry 4.0 era is coming. This causes companies and agencies are competing in building a system that integrates cyber technology and automation technology. One example is the system for recognizing, detecting, and sorting cars according to their type as the number of cars dramatically increases. For the case when calculating certain types of cars passing the road, instant and accurate sorting of car types is required. This certainly influences the development process of related agencies. This paper presents an accuracy improvement of convolutional neural network in MobileNet-V1 and ResNet-50 models using Adam setting for car type classification. Maximum results are obtained by varying different parameters, the use of the optimizer Adam and RMSprop, employing ImageNet as transfer learning, as well as implementing dropout layers and 2 kinds of learning rate. This system is trained using 10,708 pictures of City Car, Pick Up, Sedan, Sport ultra-vehicle, and Van. Accuracy of car type classification reaches 92,1% for MobileNet-V1 and 88,2% for ResNet-50 models.