Dynamic generative R-CNN

Abstract

Different techniques have been developed for object detection and recognition. These techniques can be divided into single-shot and two-shot methods. Single-shot methods focus on real-time applications, while two-shot methods are used in applications requiring higher accuracy. However, different versions of the two-shot techniques produce limited results in terms of accuracy and speed, or both. Therefore, this study proposes a novel model called dynamic generative R-CNN (DGR-CNN) that reduces the number of proposed regions using a dynamic programming model that applies the graph similarity method over graph-based image segmentation. Additionally, the proposed model employs DCGAN technique to improve detection performance. DGR-CNN reduces the overall detection and classification time and enhances the detection accuracy. The PASCAL VOC2007 and MS COCO datasets were utilized to evaluate the model. The results showed that DGR-CNN significantly reduces the number of candidate regions compared to the selective search algorithm employed in R-CNN and fast R-CNN. Although fast R-CNN utilizes 2000 regions and faster R-CNN utilizes 300 regions, DGR-CNN reduces the number of regions to approximately 130. The mean average precision of the proposed method was 75.1% on the PASCAL VOC2007, while fast and faster R-CNN scored 66.9% and 69.9%, respectively. Moreover, the DGR-CNN model significantly improved the classification accuracy when tested on the MS COCO dataset, achieving an MAP of 68.76%, compared with 32.64% and 42.3% for fast and faster R-CNN. This increase in accuracy was achieved without significantly compromising the speed compared with faster R-CNN.