A Classification Method for Unordered Steel Scrap Images Based on Hybrid Convolutional Networks

With the development of deep learning and image recognition technologies, these methods have found widespread applications across various fields. However, the annotation of industrial application-specific data, such as steel scrap images, remains challenging due to the unordered nature of the images, which often requires domain experts to perform the annotation. To address this, we propose a feature learning method for unordered images based on hybrid convolutional networks. Utilizing a preprocessed steel scrap dataset, we employ a cascaded convolutional network to capture deep correlation features within such unordered images while enhancing their fine-grained structures. A self-supervised approach is used to train the encoder, which is subsequently fine-tuned using a minimal set of labeled samples to complete the classification and annotation of steel scrap. Experimental results on four different categories of steel scrap annotation demonstrate that the proposed method achieves accuracies of 60.47%,61.24%,23.45%,62.46%, respectively, verifying the effectiveness of the method.