Feature Fusion Network Based on Academic Content Detection

At a time when academic achievements are being protected with emphasis the testing of academic contents has become a hot topic of much attention at present. In order to avoid the phenomenon of "multiple use of one image" and "image plagiarism" in the academic world, in view of the defects of insufficient feature fusion and low efficiency in the current target detection model, we propose a more powerful model method AIDPaper based on Faster R-CNN. AIDPaper adds PAFPN and DIoU Loss to the Faster R-CNN model. PAFPN is a bottom-up path added on top of FPN to greatly shorten the information propagation path and avoid information loss to a large extent. Subsequently, adaptive feature pooling is added to perform feature fusion for each regional proposal projected onto the feature layer to further enhance the fusion information of the features. DIoU loss provides the prediction direction for the prediction box making the regression more accurate not only speeding up the convergence but also reducing the detection time of the model. We perform experimental comparisons on the PaperImg dataset. The experimental results show that our proposed AIDPaper performs better and the mAP of AIDPaper is much higher than that of the baseline model Faster R-CNN at 3.2% ahead of Libra R-CNN at 6.1%. It has been shown experimentally that our model is more competitive compared to current detection models.