Multiclass Altitude and Rotation Invariant Object Detection Using Slicing Objectness Loss (SOL) For Images with Complex Background

The difficulty of detecting and monitoring diverse types of solar panels from various elevations and complex backgrounds has not been investigated. The varying orientation and complex backgrounds lead to false positives and erroneous classifications. However, by employing our technique, we could precisely distinguish between three diverse types of solar panels in various challenging backgrounds. We propose a slicing objectness loss specifically to address model learning issues, false positives, and oriented object localization on complex backgrounds. We based our work on popular existing neural network architectures and designed them to adapt to areal object detection for complex backgrounds. To our knowledge in the multi-solar panel, detection has not been conducted previously due to a lack of data availability. We have designed the neural network to process the full-size image and extract multi-Altitude local and global features thus our models are all altitude invariant and can detect shapes of arbitrary orientation. To demonstrate Slicing loss function produces superior results, we present a comparison study between pre and post-application SOL. We show clear evidence that our slicing approach and slicing objectness loss (SOL) has a significant effect on multi-solar panel detection WPV (water heater Photovoltaic), FPV (farm type photovoltaic), and SPV (Single Photovoltaic). We have also shown our approach and custom loss works for other complex multi-object detection e.g., identifying enclosures, water, and fuel tanks from different altitudes.