Research on Clothing Image Classification Models Based on CNN and Transfer Learning

Traditional clothing image classification method, which involves artificially extracting features of clothing images, are not effective owing to images having complex background and clothing distortion. With the purpose of improving the classification accuracy, eight models are designed based on convolutional neural network (CNN) and transfer learning. These CNN models are trained on the dataset named Street-FashionData having eight categories of clothing images, and their performances are evaluated and compared by conducting classification tests. The experimental results show that model ICM-C and model ECM-C have high classification accuracies of 97.2% and 97.8%, respectively. These two models have the same convolutional bases as pre-trained Inception-v3 and pre-trained EfficientNet-B0, respectively, moreover, their entire networks need to be trained. The experimental results have proved the efficiency of classification models described in this paper.