Abstract:
The present invention discloses a system and method for detecting objects in real-time adverse weather-degraded scenes. A single-stage CNN architecture is adopted, namely, AWDRDNet for detecting objects more accurately in adverse weather-degraded realistic scenes. The present invention relates to a feed-forward deeper convolutional layer comprising a plurality of convolution blocks (B1,..,BK,..,BN) producing better quality of restoration images (RI1,..,RIK,..,RIN); wherein receptive field plays an important role in analyzing local features over degraded scenes. Another key feature of the proposed invention is the clipping of pre-defined multi-scale anchor boxes per cell to a restorated de-convolutional feature map (DC) only at the top of the network, which allows to efficiently reduce time-consumption. In terms of detection accuracy (recall-precision graph and mAP), the results of the reference dataset demonstrates the optimal performance of the proposed model and reveals the performance accuracy in low-light or rainy conditions to be higher than that in dusty or foggy conditions.
|
Type: Application
Filed: January 20, 2021
Publication Date: Pending
Applicants: Anu Singha (Ph.D Scholar, Department of Computer Science and Engineering, Tripura University), Sourav Dey Roy (Ph.D Scholar, CSIR-SRF-Direct, Department of Computer Science and Engineering, Tripura University), Mrinal Kanti Bhowmik (Assistant Professor, Department of Computer Science & Engineering, Tripura University)
|