Small Target Detection Based on Squared Cross Entropy and Dense Feature Pyramid Networks

Abstract

At present, the research hotspot of small target detection mainly focuses on the methods based on deep learning, and such algorithms still have some problems that need to be solved, such as the foreground-background class imbalance and the poor performance of multi-scale detection. For the foreground-background class imbalance, the Squared Cross Entropy (SCE) loss function is proposed here to help solve the problem. Meanwhile, as Feature Pyramid Networks (FPN) is a powerful means to deal with multi-scale detection problems, a new Dense FPN structure is designed based on FPN. The Dense FPN removes the up-sampling process in FPN, and after each feature extraction layer, a continuous convolutional layer with a decreasing number of layers is added. According to the experimental results, Dense FPN outperform the original FPN on various evaluation indicators like $AP_{S}$ , $AP_{M}$ and $AP_{L}$ , showing the excellent performance of the Dense FPN in dealing with multi-scale detection problems.

Keywords

• Feature Pyramid Networks
• loss function
• target detection
• object detection