Spatial reconstruction feature for object detection on coal mine

Object detection has extensive application demands in coal mine operation scenarios. However, conducting high-quality object detection under complex underground backgrounds and harsh environmental conditions remains challenging. Most existing object detection methods adopt feature pyramids to retain the spatial and semantic information of the objects. However, it remains a critical challenge to extract features that preserve detailed object information from degraded images of coal mine workfaces. To address this issue, we propose a spatial reconstruction feature generation method for object detection from underground coal mines. The method first upsamples the feature pyramid via transposed convolutions to generate spatial reconstruction features. Subsequently, to ensure the effectiveness of the generated spatial reconstruction features, the stepwise super-resolution feature reconstruction loss supervises the feature pyramid layer by layer, introducing key target information from the image into the spatial reconstruction feature pyramid to assist the target detection process. The results in the underground coal mine target detection dataset DSLMF+ show that after introducing spatial reconstruction features, the mAP@50 of the YOLO-v5 object detection framework can be increased by an average of 1.65%, and the mAP@50 of the DETA object detection framework can be increased by an average of 2.42%. These experimental results demonstrate the effectiveness of the proposed method.