It is important to accurately classify the defects in hot rolled steel strip since the detection of defects in hot rolled steel strip is closely related to the quality of the final product. The lack of actual hot-rolled strip defect data sets currently limits further research on the classification of hot-rolled strip defects to some extent. In real production, the convolutional neural network (CNN)-based algorithm has some difficulties, for example, the algorithm is not particularly accurate in classifying some uncommon defects. Therefore, further research is needed on how to apply deep learning to the actual detection of defects on the surface of hot rolled steel strip. In this paper, we proposed a hot rolled steel strip defect dataset called Xsteel surface defect dataset (X-SDD) which contains seven typical types of hot rolled strip defects with a total of 1360 defect images. Compared with the six defect types of the commonly used NEU surface defect database (NEU-CLS), our proposed X-SDD contains more types. Then, we adopt the newly proposed RepVGG algorithm and combine it with the spatial attention (SA) mechanism to verify the effect on the X-SDD. Finally, we apply multiple algorithms to test on our proposed X-SDD to provide the corresponding benchmarks. The test results show that our algorithm achieves an accuracy of 95.10% on the testset, which exceeds other comparable algorithms by a large margin. Meanwhile, our algorithm achieves the best results in Macro-Precision, Macro-Recall and Macro-F1-score metrics.

Hot-rolled strip steel is widely used in automotive manufacturing, chemical and home appliance industries, and its surface quality has a great impact on the quality of the final product. In the manufacturing process of strip steel, due to the rolling process and many other reasons, the surface of hot rolled strip steel will inevitably produce slag, scratches and other surface defects. These defects not only affect the quality of the product, but may even lead to broken strips in the subsequent process, seriously affecting the continuation of production. Therefore, it is important to study the surface defects of strip steel and identify the types of defects in strip steel. In this paper, a scheme based on ResNet50 with the addition of FcaNet and Convolutional Block Attention Module (CBAM) is proposed for strip defect classification and validated on the X-SDD strip defect dataset. Our solution achieves a classification accuracy of 94.11%, higher than more than a dozen other compared deep learning models. Moreover, to adress the problem of low accuracy of the algorithm in classifying individual defects, we use ensemble learning to optimize. By integrating the original solution with VGG16 and SqueezeNet, the recognition rate of oxide scale of plate system defects improved by 21.05 percentage points, and the overall defect classification accuracy improved to 94.85%.

In recent years, benefiting from deep convolutional neural networks (DCNNs), face parsing has developed rapidly. However, it still has the following problems: (1) Existing state-of-the-art frameworks usually do not satisfy real-time while pursuing performance; (2) similar appearances cause incorrect pixel label assignments, especially in the boundary; (3) to promote multi-scale prediction, deep features and shallow features are used for fusion without considering the semantic gap between them. To overcome these drawbacks, we propose an effective and efficient hierarchical aggregation network called EHANet for fast and accurate face parsing. More specifically, we first propose a stage contextual attention mechanism (SCAM), which uses higher-level contextual information to re-encode the channel according to its importance. Secondly, a semantic gap compensation block (SGCB) is presented to ensure the effective aggregation of hierarchical information. Thirdly, the advantages of weighted boundary-aware loss effectively make up for the ambiguity of boundary semantics. Without any bells and whistles, combined with a lightweight backbone, we achieve outstanding results on both CelebAMask-HQ (78.19% mIoU) and Helen datasets (90.7% F1-score). Furthermore, our model can achieve 55 FPS on a single GTX 1080Ti card with 640 × 640 input and further reach over 300 FPS with a resolution of 256 × 256, which is suitable for real-world applications.

Diabetic retinopathy (DR) is a common fundus disease that leads to irreversible blindness, which plagues the working-age population. Automatic medical imaging diagnosis provides a non-invasive method to assist ophthalmologists in timely screening of suspected DR cases, which prevents its further deterioration. However, the state-of-the-art deep-learning-based methods generally have a large amount of model parameters, which makes large-scale clinical deployment a time-consuming task. Moreover, the severity of DR is associated with lesions, and it is difficult for the model to focus on these regions. In this paper, we propose a novel deep-learning technique for grading DR with only image-level supervision. Specifically, we first customize the model with the help of self-knowledge distillation to achieve a trade-off between model performance and time complexity. Secondly, CAM-Attention is used to allow the network to focus on discriminative zone, e.g., microaneurysms, soft/hard exudates, etc.. Considering that directly attaching a classifier after the Side branch will disrupt the hierarchical nature of convolutional neural networks, a Mimicking Module is employed that allows the Side branch to actively mimic the main branch structure. Extensive experiments are conducted on two benchmark datasets, with an AUC of 0.965 and an accuracy of 92.9% for the Messidor dataset and 67.96% accuracy achieved for the challenging IDRID dataset, which demonstrates the superior performance of our proposed method.

A new Vision Transformer(ViT) model is proposed for the classification of surface defects in hot rolled strip, optimizing the poor learning ability of the original Vision Transformer model on smaller datasets. Firstly, each module of ViT and its characteristics are analyzed; Secondly, inspired by the deep learning model VGGNet, the multilayer fully connected layer in VGGNet is introduced into the ViT model to increase its learning capability; Finally, by performing on the X-SDD hot-rolled steel strip surface defect dataset. The effect of the improved algorithm is verified by comparison experiments on the X-SDD hot-rolled strip steel surface defect dataset. The test results show that the improved algorithm achieves better results than the original model in terms of accuracy, recall, F1 score, etc. Among them, the accuracy of the improved algorithm on the test set is 5.64% higher than ViT-Base and 2.64% higher than ViT-Huge; the accuracy is 4.68% and 1.36% higher than both of them, respectively.