.png){kind=icon}
.png){kind=icon}
Return
Residual manganese blocks detection algorithm based on CD-SSD
.png){kind=icon}
PRE
AI
DOI：10.1088/2631-8695/ae0d42.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object detection is crucial in industrial automation, especially for tasks like automated monitoring and quality control. In the electrolytic manganese production process, identifying residual manganese blocks on cathode plates is vital for improving production efficiency and product quality. These blocks present unique challenges due to significant scale variations and shape distortions caused by adhesion, necessitating specialized detection approaches. Using traditional methods often relies on manual configurations to detect these small, irregularly shaped blocks, leading to notable accuracy declines. To address this, we propose an enhanced Single Shot MultiBox Detector (SSD) algorithm specifically designed for detecting residual manganese blocks. Our framework incorporates three key innovations: (1) Semantic Dilated Fusion (SDF), which merges semantic information from high-level features into dilated low-level features, enhancing contextual semantics while preserving fine-grained details; (2) Residual connections to strengthen original feature weights during multi-scale fusion; and (3) a Concatenation-Driven Gated Channel Transformation (CDGCT) mechanism that balances channel-wise dependencies by processing concatenated features. Extensive experiments on a proprietary industrial dataset and the PASCAL VOC benchmark demonstrate the model's efficacy, achieving mean average precision (mAP) values of 84.63% and 82.30%, respectively, surpassing state-of-the-art methods such as YOLOv11, Faster R-CNN, DSSD, and others. This work offers a robust and generalizable solution for high-precision detection in industrial automation scenarios.
Keywords:
residual manganese blocks detection
single shot multibox detector
semantic dilated fusion
concatenation-driven gated channel transformation
Journal
E
IF:
1.6
Papers:
2.1K
Citations:
0