Hong Wang

In the maritime industry, accurately detecting a ship’s draft line is crucial for ensuring transaction fairness and navigational safety. Existing deep learning-based methods for draft line detection primarily use segmentation techniques to segment the entire body of water before determining the waterline. These approaches incur high computational costs and often face challenges under varying environmental conditions, such as lighting changes and different hull colors. To address these issues, we propose multi-scale feature fusion keypoint detection network (MFFKD) for precise and efficient ship draft line detection. Our network integrates four stages of Dilated Residual-Channel Recalibration Module (DR-CRM) blocks to extract multi-scale features. Meanwhile, the Feature Enhancement Extraction Modules (FEEM) are employed to enhance these extracted features, and the Multi-scale Feature Weighted Integration (MFWI) module efficiently fuses the enhanced multi-scale features. Furthermore, a task head for keypoint prediction is designed to ensure accurate localization of keypoints. By integrating the predicted keypoint data with mark information detected by the character recognition head through a mathematical model, we achieve precise predictions of waterline readings. To enhance the model’s adaptability to various environmental conditions, we adopt a dual-phase training strategy: an initial pre-training phase for learning general ship features and waterline characteristics, followed by a fine-tuning phase using data from diverse scenes. Extensive experimental results show that our method surpasses the baseline models in waterline detection accuracy. In terms of model execution speed, our method exceeds the advanced segmentation-based approaches. These demonstrate the effectiveness of integrating keypoint detection with dual-phase training in ship waterline detection.

Accurate measurement of the position and shape of the blast furnace (BF) burden surface (BS) is crucial for automated and intelligent BF control. The detection method based on key points of BS can improve the accuracy of radar BS detection in challenging environments with strong interference, ensuring long-term detection accuracy stability across long maintenance intervals. However, transferring a successfully applied detection model from the original BF to a new target BF requires sufficient sample collection in the target BF and model retraining. In practical scenario, data collection is often constrained by plenty of factors like time, space, and personnel. Thus, reducing the number of samples needed for retraining, or implementing model transfer based on small samples, can significantly boost BS detection efficiency. This paper proposes a Conv-Adapter Transfer Net (CATNet) designed for the transfer problem of key point detection models in BFBS and small sample scenarios. The CATNet adopts a newly designed parallel Conv-Adapter transfer module, a small-scale deep learning network with a minimal parameter quantity. Compared to fine-tuning the entire backbone model with a large number of parameters, retraining this transfer module requires far less training data. Therefore, CATNet is more suitable for model transfer in small sample scenarios. The experimental results demonstrate that the CATNet can exhibit strong transfer capability and higher prediction accuracy than the commonly used pretrain-finetune method in small sample scenarios, thereby enhancing the efficiency of model transfer across different BFs.

Semantic segmentation by using remote sensing images is an efficient method for agricultural crop classification. Recent solutions in crop segmentation are mainly deep-learning-based methods, including two mainstream architectures: Convolutional Neural Networks (CNNs) and Transformer. However, these two architectures are not sufficiently good for the crop segmentation task due to the following three reasons. First, the ultra-high-resolution images need to be cut into small patches before processing, which leads to the incomplete structure of different categories’ edges. Second, because of the deficiency of global information, categories inside the crop field may be wrongly classified. Third, to restore complete images, the patches need to be spliced together, causing the edge artifacts and small misclassified objects and holes. Therefore, we proposed a novel architecture named the Coupled CNN and Transformer Network (CCTNet), which combines the local details (e.g., edge and texture) by the CNN and global context by Transformer to cope with the aforementioned problems. In particular, two modules, namely the Light Adaptive Fusion Module (LAFM) and the Coupled Attention Fusion Module (CAFM), are also designed to efficiently fuse these advantages. Meanwhile, three effective methods named Overlapping Sliding Window (OSW), Testing Time Augmentation (TTA), and Post-Processing (PP) are proposed to remove small objects and holes embedded in the inference stage and restore complete images. The experimental results evaluated on the Barley Remote Sensing Dataset present that the CCTNet outperformed the single CNN or Transformer methods, achieving 72.97% mean Intersection over Union (mIoU) scores. As a consequence, it is believed that the proposed CCTNet can be a competitive method for crop segmentation by remote sensing images.

Accurate Burden Surface Profile (BSP) detection is important for the operation of Blast Furnace (BF). The signal-to-noise ratio of radar signals changes greatly during both the charging period of BF and the long maintenance period of the radar device, which increases the difficulty of radar BSP detection. The traditional radar BSP detection method based on signal energy relies on manually selected detection thresholds according to the noise intensity. Hence, the accuracy of the traditional radar BSP detection method is not reliable in the long term. To address this problem, we propose a novel learning-based Key Points estimation (KP-BSP) method to detect the key points of radar reconstructed BSP image, and a new Key Points-based Connected Region Noise Reduction (KP-CRNR) algorithm to remove the noise-affected regions. The prediction deviation at detected key points (at the positions of the mechanical probes) is then used to correct the radar detection results, leading to the improvement of radar detection accuracy. The experimental data were collected from Wuhan Iron Steel Company No.7 BF. The results show that the proposed methods can achieve an average RMSE of 0.0156m, which is improved by more than 50% compared with previous methods. The long-term reliability of the proposed method is also demonstrated in this dataset.