Geological Image Defect Processing Based on ResNet50 with Transfer Learning in Visual Communication

In geological research, classification of rock image is a significant role. However, existing deep learning (DL) techniques rock classification accuracies are unsatisfied because of network model's weak feature extraction. This paper implements a ResNet50 model that uses the transfer learning (TL) approach to develop an automatic classification model for rock image classification. ResNet50 creates the residual structure to enhance its performance in image classification by extracting high-quality features from rock images. When there are fewer rock images, the TL approach better express the rock texture information for classification by abstracting the deep features from shallow features. Rock and ImageNet datasets are used to analyze this model. The implemented method outperformed better performance than existing methods such as ResNet34, Triplet-EfficientNet, and InceptionResnetV2, in terms of 95.90% accuracy using rock dataset, Compared to You Only Look Once (YOLO) version 4, implemented ResNet50 achieved 99.99% accuracy, 99.99% precision, 99.90% recall, and 99.95% f-measure using ImageNet dataset.