Optimized Scale-Invariant Hog Descriptors for Tobacco Plant Detection

Abstract

The histogram of gradient (HOG) descriptor is being employed in this research work to demonstrate the technique of scale variant to identify the plant in surveillance videos. In few scenarios, the discrepancies in the histogram of gradient descriptors along with scale as well as variation in illumination are considered as one of the major hindrances. This research work introduces a unique SIO-HOG descriptor that is approximated to be scale-invariant. With the help of the footage that is captured from the tobacco plant identification process, the system can integrate adoptive bin selections as well as sample resizing. Further, this research work explores the impact of a PCA transform that is based on the process of feature selection on the performance of overall recognition and thereby considering finite scale range, adoptive orientation binning in non-overlapping descriptors, as well as finite scale range are all essential for a high detection rate. The feature vector of HOG over a complete search window is computationally intensive. However, suitable frameworks for classification can be developed by maintaining a precise range of attributes with finite Euclidean distance. Experimental results prove that the proposed approach for detecting tobacco from other weeds has resulted in an improved detection rate. And finally, the robustness of the complete plant detection system was evaluated on a video sequence with different non-linearity's that is quite common in a real-world environment and its performance metrics are evaluated