MMP-YOLO: A multi-branch defect detection model based on rich gradient information

In the manufacturing process, the diversity of products and the complexity of the production environment pose severe challenges to the detection of small defects, which can lead to serious missed detections and false positives. To address these issues, this paper proposes a multi-branch defect detection model based on rich gradient information (MMP-YOLO), which significantly improves the performance of detecting defective objects. Specifically, we design three innovative modules integrated into MMP-YOLO. (1) The Multi-Level Gradient Lightweight Deep Network (MGLD) module processes multi-gradient information through a deep network integrated with large kernel convolution, ensuring accurate transmission of original input information and efficient feature extraction of small objects. (2) The Multi-Scale Function Complementary Upsampling (MFCU) module exploits the complementarity between high-resolution and low-resolution features and introduces transposed convolution and dilated convolution to reduce information loss further. (3) The Parallel Task-Related Feature Selection (PTFS) module selectively suppresses background interference through a combination of global and local information. Extensive experiments on multiple datasets demonstrate that MMP-YOLO outperforms other state-of-the-art methods in reducing information loss and minimizing background noise interference. • We propose a novel feature extraction module (MGLD), which minimizes information loss during gradient flow propagation. • We designed the Multi-Scale Feature Complementary Upsampling (MFCU) module that uses different feature fusion methods on feature maps of different resolutions. • We develop the Parallel Task-Related Feature Selection (PTFS) module to separate the foreground and background of the block-processed feature maps. • By integrating the three novel modules described above, MMP-YOLO demonstrates superior performance on multiple defect detection datasets.