New Feature Method Boosts Coal Mine Object Detection

Researchers from China Coal Technology & Engineering Group and affiliated institutions have developed a spatial reconstruction feature generation method designed to improve object detection in the challenging environments of underground coal mines. Published in Frontiers in Artificial Intelligence in October 2026, the study addresses the persistent difficulty of identifying objects in degraded images characterized by non-uniform lighting, dust, and complex backgrounds. The proposed method integrates super-resolution information into the object detection task by reconstructing feature pyramids, aiming to preserve detailed object information that is typically lost during standard downsampling operations in convolutional neural networks.

The core challenge in underground coal mine workfaces is that complex background features and non-object elements can interfere with the discriminative features of small objects. While existing object detection methods often adopt feature pyramids to retain spatial and semantic information, these structures are insufficient for the harsh conditions found in mining environments. The authors propose a module that upsamples the feature pyramid via transposed convolutions to generate spatial reconstruction features. This process is supervised by a stepwise super-resolution feature reconstruction loss, which ensures that the generated features contain effective object information by introducing key target details from the original image into the feature pyramid.

The method operates in two primary steps: upsampling and learning the reconstructed features. First, a set of learnable upsampling modules unifies the feature sizes of different pyramid levels. This involves two layers of 3×3 convolution to increase the receptive field, followed by a 3×3 transposed convolution that upsamples features through learnable parameters. A 1×1 convolutional layer then performs channel transformation.

To ensure the quality of the reconstructed features, the authors introduce a ladder super-resolution reconstruction loss. This loss function supervises the feature pyramid layer by layer, fitting the appearance differences between the upsampled features and the real image. The loss incorporates an adjacency subgraph constructed from pixel features, where each pixel feature constructs an adjacency matrix based on K-Nearest Neighbor nodes. This approach leverages neighborhood information to enhance the discriminative power of pixel features, allowing the model to better distinguish target-related characteristics, such as reflective strips on miner clothing, from background clutter.

The study evaluated the proposed method on two large-scale datasets: DSLMF+ and CUMDS. The DSLMF+ dataset, designed for underground long-wall fully mechanized mining workspaces, contains 138,004 images labeled with six types of objects, including miners, hydraulic support guard plates, large coal blocks, trailing cables, miner behaviors, and mining safety helmets. The CUMDS dataset, focused on underground drilling sites, includes 70,948 images covering five categories: gripper, chuck, coal miner, mine safety helmet, and drill pipe. These datasets provide a robust benchmark for testing the method’s effectiveness in diverse and degraded mining environments.

Experimental results on the DSLMF+ dataset demonstrate significant improvements in detection performance. When the spatial reconstruction features were introduced into the YOLO-v5 object detection framework, the mean average precision at 50% Intersection over Union (mAP@50) increased by an average of 1.65%, and the mAP@50:95 increased by 2.03%. For the YOLO-v7 framework, the mean mAP@50 improved by 0.97%, while the mAP@50:95 saw a more substantial gain of 3.85%. These results indicate that the method is particularly effective at enhancing precise object localization, as the larger improvements in mAP@50:95 suggest that the reconstructed features help the model achieve higher Intersection over Union scores under strict localization thresholds.

The method was also tested on Transformer-based frameworks, including ViT-Adapter-L and DETA. For the ViT-Adapter-L framework, the spatial reconstruction feature generation module was added to each Spatial Prior Module, resulting in mean increases of 0.78% in mAP@50 and 3.40% in mAP@50:95. In the DETA framework, which utilizes a one-to-many sample-matching strategy, the incorporation of spatial reconstruction features into the CNN backbone increased the mean mAP@50 from 93.53% to 95.95%, and the mAP@50:95 from 67.37% to 71.70%. These findings confirm that the proposed features are generalizable across both CNN-based and Transformer-based detection architectures.

Performance on the CUMDS dataset further highlights the method’s utility in degraded environments. On this dataset, the spatial reconstruction features boosted the mean mAP@50 and mAP@50:95 by 0.7% and 6.9%, respectively, for the YOLO-v5 framework. For YOLO-v7, the improvements were 0.5% and 5.4%. The Transformer-based frameworks also showed notable gains, with ViT-Adapter-L improving by 1.8% in mAP@50 and 4.8% in mAP@50:95, and DETA increasing its mAP@50:95 by 2.7%. The larger gains observed on CUMDS compared to DSLMF+ suggest that the method is especially effective in environments with severe image degradation, such as those affected by dust, low illumination, and motion blur.

Ablation studies conducted on the DSLMF+ sub-datasets for coal miners and mine safety helmets isolated the contribution of each component. The base YOLO-v5 model achieved an mAP@50 of 95.1% for coal miners. Incorporating only the pixel loss for reconstruction features slightly reduced this to 94.9%, but adding the adjacency loss improved the mAP@50 to 96.0% and the mAP@50:95 to 74.1%. The full method, combining both upsampling and loss components, achieved the highest performance with an mAP@50 of 96.7% and an mAP@50:95 of 74.3%. For mine safety helmets, the full method improved the mAP@50 from 94.8% to 98.0% and the mAP@50:95 from 65.6% to 71.4%, demonstrating the significant impact of the spatial reconstruction features on small object detection.

The computational overhead of the proposed method was also analyzed. Incorporating the spatial reconstruction feature module increases the number of parameters, FLOPs, and GPU memory requirements. Similar increases are observed for other frameworks, with the DETA framework seeing its parameters rise from 210.0 million to 237.2 million. These additional resources are used to reconstruct features and introduce spatial characteristic information, which the authors argue is necessary to compensate for the losses in spatial cues caused by image degradation.

Visualization results provide intuitive evidence of the method’s effectiveness. In images from the DSLMF+ and CUMDS datasets, the base YOLO-v5 model often struggles to distinguish objects from the background clutter, leading to missed detections. After incorporating the spatial reconstruction features, the model better distinguishes objects from the background, reducing missed detections. The reconstructed object-pixel features enable the detection model to leverage neighborhood information, enhancing the robustness of pixel features against target-irrelevant features. This is particularly evident in detecting miners, where reflective strips on clothing serve as key discriminative cues that are better preserved and utilized by the enhanced feature pyramid.

The study acknowledges that degraded images from underground coal mine environments continue to adversely affect object detection performance. The proposed method includes more learnable parameters to reconstruct features and reduce these effects, which requires more computational resources. The authors note that future work will focus on integrating image denoising with object detection frameworks to develop an end-to-end approach for joint image denoising and object detection.

In conclusion, the spatial reconstruction feature generation method offers a flexible and effective plug-and-play component for both CNN-based and Transformer-based object detection frameworks. By forcing the feature pyramid to retain fine-grained spatial details that are typically lost during downsampling, the method enhances the model’s ability to detect small objects in complex and degraded environments. The consistent improvements across multiple datasets and frameworks validate the effectiveness of the proposed approach in addressing the specific challenges of underground coal mine workfaces.

Subject of Research: Spatial reconstruction feature for object detection on coal mine

Article Title: Spatial reconstruction feature for object detection on coal mine

Article References: Ji, H., Jiang, P., Wang, Z., Chen, Z., & Liu, Y. (2026). Spatial reconstruction feature for object detection on coal mine. Frontiers in Artificial Intelligence, 9, Article 1959087. https://doi.org/10.3389/frai.2026.1959087

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1959087

Keywords: Spatial, reconstruction, feature, object, detection, coal, mine, scientific research