GPCNN: Grouped Pooling Convolutional Neural Network for Radar Target Detection

Radar target detection plays an important role in radar signal processing. Constant False Alarm Rate (CFAR) is the most widely used detection method, while its performance is easily affected by clutter and noise and greatly degrades in the condition with multiple targets. In this paper, we propose a Grouped Pooling Convolutional Neural Network (GPCNN) model to address the aforementioned issues. On one hand, this model extracts the deep features to detect targets from complex background. On the other hand, multiple tricks, including pooling module, replacement of fully connected layers, shuffle concatenation, and weight transfer, are introduced into CNN to improve the detection efficiency. Experimental results on the measured data demonstrate that the target detection probability of our proposed GPCNN model has increased by 10% compared with CFAR, and the operation speed has increased by 12 times compared with traditional CNN model.