Deep-NFA: a Deep $\textit{a contrario}$ Framework for Small Object Detection

TL;DR

This work introduces an $\textit{a contrario}$ decision criterion into the learning process to take into account the unexpectedness of small objects and enhances the feature map responses while controlling the number of false alarms (NFA).

Abstract

The detection of small objects is a challenging task in computer vision. Conventional object detection methods have difficulty in finding the balance between high detection and low false alarm rates. In the literature, some methods have addressed this issue by enhancing the feature map responses, but without guaranteeing robustness with respect to the number of false alarms induced by background elements. To tackle this problem, we introduce an $\textit{a contrario}$ decision criterion into the learning process to take into account the unexpectedness of small objects. This statistic criterion enhances the feature map responses while controlling the number of false alarms (NFA) and can be integrated into any semantic segmentation neural network. Our add-on NFA module not only allows us to obtain competitive results for small target and crack detection tasks respectively, but also leads to more robust and interpretable results.

Figures (8)

• Figure 1: NFA and significance values for a centered unit variance Gaussian variable, with $N_{test}=1$.
• Figure 2: Diagram showing the integration of our NFA module into a U-shaped segmentation NN. Optional blocks are drawn in dotted lines. Details for ECA block can be found in the original paper eca_net.
• Figure 3: Diagram of (a) the basic NFA block and (b) the spatial NFA block. The details of the stand-alone self-attention (SASA) block can be found in ramachandran2019stand.
• Figure 4: Variations of $\textsc{Sigm}_{\alpha}$ function defined in \ref{['eq:sigm_a']}, with different values of $\alpha$. For simplicity, we choose $N_{test}=1$.
• Figure 5: Qualitative results obtained with different detection methods (columns (c) to (e)) on NUAA-SIRST dataset. Good detections and false positives are circled in red and green lines respectively, and missed detections in light yellow dotted lines. The ground truth is circled in green dotted lines in column (a). A zoom on the targets is displayed in the top right corner.