Abstract: Accurately capturing the burden surface profile information of a blast furnace is crucial for adjusting the burden distribution matrix and improving gas flow distribution, which is essential in the steel smelting industry. However, during ironmaking operations, the burden surface profile morphology exhibits dynamic stochastic roughness and manifests distinct multiphase regimes, including bubbling, fluidized, and spouting states. Traditional single-state burden surface profile detection neural network disable to account for alternating smelting states in the complex environment of a blast furnace. In this paper, a detection network architecture for multi-state burden surface profile images called BurdenNet is proposed. It is based on prior information and aims to solve the problem of low overall accuracy of the single-state burden surface profile detection neural network in complex environments. First, the range precision of the original radar signal combined with the signal-to-noise ratio and phase noise was defined. This served as a criterion for pre-classification by constructing three typical state burden surface profile datasets. The network structure was pruned using the classification results as prior information to enhance the detection rate. Second, based on the low-curvature shape feature of slender burden surface profile targets and the sparse properties of the signal sample, an Atrous Vertical Deformable Convolution (AVDC) module was proposed to extract multi-state burden surface profile features. The convolution kernel integrates both dilated and deformable convolutions, requiring only vertical offset computation. In addition, mechanical probe data from the blast furnace were utilized to construct a prior spatial attention feature map. A Prior Focusing Attention (PFA) module was proposed, leveraging this map for spatial feature extraction, allowing the network to focus more effectively on the burden surface profile region in the image. Finally, a Band Intersection Over Union (BIOU) loss function was proposed for anchor-free boundary box regression, further improving detection accuracy and speed by eliminating X-coordinate computations from the BIOU calculation. Experimental results on measured data from the blast furnaces in iron and steel companies demonstrate that, compared with traditional burden surface profile detection networks, the proposed BurdenNet improves detection accuracy by 13.9% and 5.2%, and enhances comprehensive performance (F1-Score) by 8.08% and 4.30%, respectively. Ablation experiments show that the proposed AVDC module improves detection accuracy by 17.7% and the F1-score by 15.6% compared with conventional convolution kernels. The proposed PFA module improves accuracy by 4.3% and the F1-score by 4.7% compared with shuffle attention (SA) and achieves a 3.9% higher accuracy and a 4.2% higher F1-score compared with Non-Local Attention (NLA). The proposed BIOU function showed a 1.7% better accuracy and a 1.1% better F1-score than the traditional CIOU function, and the detection FPS was improved by 10.4%. These results demonstrate that BurdenNet provides a more accurate method for the detection of multi-state burden surface profile microwave images in complex and confined environments. Moreover, within the attention module, the burden surface profile detection task places particular emphasis on the spatial characteristics of the extracted features.