Abstract: Constellation diagram is a widely used feature expression method in automatic modulation classification technology. At present, a variety of constellation diagram enhanced methods have been developed to improve its representation effectiveness. However, most of the methods can not accurately extract the constellation point density in different constellation regions, and the contrast enhancement effect on constellation clusters with different densities is limited. In order to overcome the above defects, this paper proposes a regional pixel density constellation diagram enhanced method. By calculating the grayscale pixel density in each small grid region, our method more accurately reflects the density of different constellation clusters. Furthermore, the pixel value is reset by the segmented pixel threshold, which greatly improves the visual contrast effect of the enhanced constellation diagram. In addition, this paper designs a complementary Deep Feature Aggregation Attention Network (DFAAN), which significantly enhances the robustness of modulation type classification by aggregating semantic features from different levels. The experimental results show that the proposed method outperforms other approaches on the dataset with 23 modulation types, even under phase offset, frequency offset, and low SNR conditions. Specifically, the OA is at least 11.26% higher, the AA is at least 8.65% higher, and the Kappa coefficient is at least 11.77% higher compared to other methods.