Abstract: Accurately capturing the burden surface profile information of a blast furnace is helpful to adjust the burden distribution matrix and improve the 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. The traditional burden surface profile detection neural network ignored the alternating change of smelting state under the complex multi-state environment in blast furnace. In this paper, a detection network architecture for multi-state burden surface profile images BurdenNet based on prior information is proposed, aiming to solve the problem that the accuracy of single fixed mode for burden surface profile image detection method is low. Firstly, a novel definition of range precision of original radar signal combined with signal-to-noise ratio and phase noise is proposed, which serves as criterion for pre-classification, constructing three typical state burden surface profile datasets. The network structure is pruned with the classification results as the prior information in order to enhance detection rate. Secondly, according to the low curvature shape feature of the slender burden surface profile target and the sparse property of the signal sample as prior information, the Atrous Vertical Deformable Convolution (AVDC) module is proposed, the convolution kernel integrates both dilated convolution and deformable convolution, while requiring only vertical offset computation. In addition, the mechanical probe data of blast furnace is considered to construct prior spatial attention feature map, and a Prior Focusing Attention (PFA) module is proposed utilizing constructed prior spatial attention feature map for spatial features extraction. Finally, the Band Intersection Over Union (BIOU) loss function is proposed for the anchor-free regression of boundary box, further improving the accuracy and speed of detection. In the calculation of the BIOU function, X-coordinate computation can be eliminated. The experimental results on the measured data from the blast furnace of iron and steel company demonstrate that, compared with the traditional burden surface profile detection network, the accuracy of the proposed BurdenNet is increased by 13.9% and 5.2%, and the comprehensive performance (F1 Score) is increased by 8.08% and 4.30%. From the results of ablation experiments, the proposed AVDC module demonstrates 17.7% absolute improvement in accuracy and 15.6% absolute improvement in F1 score compared with conventional?convolution kernel. The proposed PFA module demonstrates 4.3% absolute improvement in accuracy and 4.7% absolute improvement in F1 score compared with shuffle attention (SA), as to non-local attention (NLA), the accuracy is 3.9% higher and F1 score is 4.2% higher. The proposed BIOU function shows 1.7% better accuracy and 1.1% better F1 score than traditional CIOU function, the detection FPS is improved by 10.4%. These provide a more accurate method for target detection in microwave images under confined and complex environment. Moreover, in the attention module, burden surface detection task places particular emphasis on the spatial characteristics of features. The network pruning is able to enhance the detection rate, but the enhancement effect is dynamically adaptive, contingent upon the holistic image quality metrics of the dataset.