研究目的
Investigating the optimal approach to deep convolutional architecture for the application of image recognition.
研究成果
Inception-ResNet was found to have the highest accuracy while keeping moderate computation requirements. Transfer learning method efficiently reduced training time as well as overfitting without affecting accuracy of the model.
研究不足
The study acknowledges the challenge of selecting the perfect strategy for implementing deep learning architecture as accuracy and time required for training depend on it. It also notes the problem of overfitting with small datasets when using models with a higher number of hidden layers.
