Severe weather phenomena are often associated with atmospheric fronts, making their identification crucial for weather analysis and forecasting. However, current machine learning methods for identifying atmospheric fronts face challenges, including a mismatch between the input data and the network, which increases training difficulty and affects accuracy. Additionally, these methods only provide category information of grid points without specifying the front they belong to, causing ambiguity when similar fronts are close together. In order to address these deficiencies, this paper proposes a Cage R-CNN architecture based on a target detection network for automatically identifying atmospheric fronts. The proposed method introduces a fusion module before the R-CNN network, which reduces conflicts between multi-meteorological factors by adjusting their weights, enhancing the identification of weak frontal features. Experimental results demonstrate that adding the fusion module to the R-CNN network improves overall performance by about 6%, making the target detection network more effective for front identification tasks. This advancement enhances the generalizability of machine learning applications to other specific tasks.