Depth Fine-Grained Recognition Method for Rock Images

In the ore mining process, rock identification is a key technology to ensure project safety. Relying on empirical methods for identification is difficult to meet real-time and accurate requirements. Rock classification tasks often face challenges such as scarcity of sample data and diversification of external features such as texture and color. To solve this problem, we propose a deep fine-grained recognition method. In order to solve the problem of small sample size, this method utilizes large-scale datasets from general image classification tasks to pre-train the model parameters. This transfer learning strategy is crucial to overcome model overfitting. Aiming at the problem of high inter-class similarities and intra-class differences caused by different rock appearances, our method enhances the model's ability to focus on rock feature information through an improved CBAM attention mechanism in the model structure design. During model training, we adopt feature-level center loss to regularize the traditional cross-entropy loss, thereby enhancing intra-class similarity and inter-class dissimilarity. Experiments show that the recognition performance of deep convolutional neural networks has been significantly improved in this task, achieving an accuracy improvement of 3.2% on the rock classification dataset.