Table 1 The research advancements in the field of sculpture image classification
From: The analysis of sculpture image classification in utilization of 3D reconstruction under K-means++
Author/year | Specific objective | Algorithm used | Advantages | Disadvantages | Basic assumption | Overall evaluation |
|---|---|---|---|---|---|---|
Nawaz et al. (2022) | Unified image representation learning and image synthesis tasks | Representation learning method based on image semantic masks | Implemented multi-task learning and demonstrated state-of-the-art performance across multiple datasets | Did not directly optimize for sculpture image classification | Image representation and synthesis could mutually enhance each other | This study provided a new perspective on image representation learning, but its direct contribution to sculpture image classification was limited |
Sujithra et al. (2024) | Image classification based on self-supervised learning | Pre-training with unlabeled image data, and fine-tuning with limited labeled data | Effectively utilized unlabeled data to improve classification efficiency | Did not fully explore the application in sculpture image classification | Self-supervised learning could learn general features | This method demonstrated the potential of self-supervised learning in image classification but required further validation for its effectiveness in sculpture image classification |
Zhai et al. (2024) | Reviewed the applications and developments of CNN in computer vision | ResNet50 | It solved the gradient vanishing problem in deep networks by introducing residual blocks | It had a large number of parameters and high computational resource requirements | Deep networks were prone to the gradient vanishing problem | ResNet50, with its innovative residual block design, significantly improved the training efficiency and performance of deep networks |
Xu et al. (2024) | Proposed a deep image classification model Based on prior feature knowledge embedding | Inception v3 | It captured multi-scale information by parallelly using convolution kernels of different sizes in each convolutional layer | Its structure was relatively complex and training was more challenging | Multi-scale information was crucial for image classification | Inception v3 enhanced the model’s expressive power and classification performance by designing multi-scale convolutional kernels |
