Instance segmentation based precise object detection in UAV Images using Mask R-CNN
Keywords:Deep learning, Instance Segmentation, Mask R- CNN, UAV Images, Optimization algorithm
Object detection plays a vital role in remote-sensing dataset which trains the image or things and helps in classifying the images into its classes. Instance segmentation is the avant-garde technique used for object detection in Deep Learning. There are number of instance segmentation models which can produce significant results. Object detection, segmentation and RGB analysis in images taken from Unmanned Aerial Vehicle (UAV) are difficult with desired level of performance. Instance segmentation is a powerful method that extracts each object and its location with the predicted label for pixels in the input image. In this paper, a study has been carried out on implementation of Mask R-CNN for instance segmentation with different optimization algorithms to obtain a more accurate result for UAV images. The training has been carried out with Mask R-CNN for object detection using ResNet50 and ResNet101 as backbone. After extensive experiments, it has been observed that the optimization algorithm plays the vital role on the overall computational process and can improve the accuracy level with reduction in the training/ validation loss. The experiment has been conducted on the publicly available UAV datasets. The paper further presents the results in terms of different performance parameters.
LicenseCopyright in a work is a bundle of rights. IJoSI's, copyright covers what may be done with the work in terms of making copies, making derivative works, abstracting parts of it for citation or quotation elsewhere and so on. IJoSI requires authors to sign over rights when their article is ready for publication so that the publisher from then on owns the work. Until that point, all rights belong to the creator(s) of the work. The format of IJoSI copy right form can be found at the IJoSI web site.
The issues of International Journal of Systematic Innovation (IJoSI) are published in electronic format and in print. Our website, journal papers, and manuscripts etc. are stored on one server. Readers can have free online access to our journal papers. Authors transfer copyright to the publisher as part of a journal publishing agreement, but have the right to:
1. Share their article for personal use, internal institutional use and scholarly sharing purposes, with a DOI link to the version of record on our server.
2. Retain patent, trademark and other intellectual property rights (including research data).
3. Proper attribution and credit for the published work.