PROJECT TITLE :

Bioinspired Scene Classification by Deep Active Learning With Remote Sensing Applications

ABSTRACT:

Scene parsing, robot motion planning, and autonomous driving are all examples of applications that require a technique that can accurately classify scenes that have a variety of spatial configurations. This technique is an essential component of computer vision and intelligent systems. Over the course of the last decade, deep learning models have accomplished remarkable feats in terms of their performance. However, to the best of our knowledge, these deep architectures are not capable of explicitly encoding the human visual perception, which refers to the sequence of gaze movements and the cognitive processes that follow. In this article, a biologically inspired deep model is proposed for the purpose of scene classification. Within this model, the human gaze behaviors are robustly discovered and represented by a unified deep active learning (UDAL) framework. More specifically, an objectness measure is used to decompose each scenery into a set of semantically aware object patches in order to characterize the components of objects that have varying sizes. This is done in order to characterize the objects' components. A local–global feature fusion scheme has been developed. This scheme optimally integrates multimodal features by automatically calculating the weight associated with each feature. The goal of this scheme is to represent each region at a low level. We developed the UDAL in order to simulate the human visual perception of a variety of different landscapes. The UDAL hierarchically represents the human gaze behavior by identifying semantically important regions within the landscape that are being viewed. Importantly, UDAL combines the detection of semantically salient regions with the deep gaze shifting path (GSP) representation learning into a principled framework. This means that all that is required are partial semantic tags. In the meantime, the incorporation of the sparsity penalty allows for the intelligent avoidance of contaminated or redundant low-level regional features. When everything is said and done, the deep GSP features that were learned from all of the scene images are combined to create an image kernel machine. This machine is then fed into a kernel SVM, which is used to categorize the various sceneries. Our method was shown to be competitive through a series of experimental evaluations on six different well-known scenery sets, which included the use of remote sensing images.