|
Original research EVALUATION OF TRANSFER LEARNING TECHNIQUE ON COMPUTER PART IMAGES CLASSIFICATIONPages 205-212 Abstract
This study makes an evaluation of the transfer learning technique applied to the classification of computer part images using three state-of-the-art convolutional neural network architectures: EfficientNetB7, ResNet50, and Xception. The study utilizes a dataset of 3,279 images with 14 distinct hardware components. Each pre-trained model was fine-tuned and evaluated under consistent experimental conditions over 10 epochs with an input resolution of 256x256 pixels. The results demonstrate the strong effectiveness of transfer learning in this specialized domain, with all models achieving high classification accuracy. EfficientNetB7 attained the highest performance at 78%, while ResNet50 reached 73%, and Xception reached 72%. This results show the most suitable model, namely EfficientNetB7 to optimize accuracy and efficiency in this problem. Meanwhile, Xception, with its depthwise separable convolutions, and ResNet50, with its residual learning framework, also delivered robust and competitive results, each showcasing distinct advantages in terms of feature extraction capability and architectural efficiency.
Keywords: Transfer Learning, Image Classification, PC Hardware, EfficientNetB7, ResNet50, Xception, Deep Learning, Computer Vision.
|
Nam Tran Quy 
PDF 