A Deep Neural Network Object Detection Method Using Multiscale Poisson Fusion

The deep neural network method has been widely used in the field of video object detection. However, the detection performance of such detectors are very sensitive to scaling change of targets. To improve detecting performance, it’s necessary to use a large number of sample data at different scales to train the model, but generally we could not get real multi-scale data of one scene. Aiming at Faster-RCNN, a state-of-the-art object detector, this paper proposes a deep neural network object detection method based on multi-scale Poisson fusion, its augment samples using Poisson method, which extracts sample objects from the original video frames and uses the Laplacian differential operator to calculate the divergence of sample objects according to the divergence of target area. After multi scaling at the sample objects, the Poisson equation is used to merge the scaling objects into the background at the target area, thus generating a large number of sample data at different scales. The experiment proves that the multi-scale sample data which generated by Possion fusion method can effectively perform the fine-tuning training for Faster-RCNN and improve its performance when detecting objects at different scales.