| Popis: |
Addressing the current industrial methods for surface defect detection, which suffer from issues such as low detection efficiency, elevated rates of false positives, and inadequate real-time capabilities, this paper proposes an high-precision industrial defect detection network. Firstly, this paper proposes a lightweight feature extraction network, which ensures real-time model detection under the premise of fully extracting defect features. Secondly, in order to solve the problem that tiny targets in industrial datasets have fuzzy texture and contain few features that are difficult to be detected, a Context Enhancement Module (CEM) is proposed, which effectively complements the contextual information of the small targets and performs multi-scale fusion to enhance the semantic information representation. Meanwhile, a Feature Enhancement Module (FEM) is designed at the end of the backbone network to optimise the feature information, effectively capture the global feature information and local feature information, and the Hybrid Attention Module (HAM) designed in this paper is introduced to extract the important information and weaken the irrelevant information in the feature map. Finally, a Dense cross-layer Feature Pyramid Network (DFPN) is proposed, which fully integrates the semantic and fine-grained features extracted from the backbone. Improve the detection of targets with significant scale changes, and adds a feature refinement module to suppress the conflicting information and reduce the semantic discrepancies before passing the learned multi-scale feature information into the prediction layer. The surface of the experimental results,the mAP@.5 was obtained at 94.3%, 98.6% and 98.4% on the steel, PCB and aluminium surface defect datasets. |
