Investigation of the performance of deep learning-based object detection systems

In recent years, significant advancements have been achieved in object detection (OD) within computer vision due to rapid developments in deep learning (DL) methodologies. Nevertheless, region-based convolutional neural networks (R-CNNs) often suffer from computational inefficiencies due to the generation of an excessive number of candidate regions, many of which are redundant or irrelevant. Although single-stage algorithms have accelerated OD, the presence of high-frequency noise and irrelevant details makes these models sensitive to background disturbances. To adress these limitations, this study introduces a novel approach reducing the high-frequency noise in the input image for the object detection task. Specifically, the proposed model comprises a novel Adaptive Padding (AP) mechanism and Region of Interest (RoI) detector, which provides a balance between RoI generation and object detection. Experimental evaluations conducted on five bird datasets using R-CNN demonstrate that our method increased the proportion of region proposals with an Intersection over Union (IoU) greater than 0,5 from 20,93% to 65,17%. Furthermore, the number of positive proposals increased by 330% during training and 726% during testing, while redundant proposals were reduced by 55,48%. The proposed model was also used for YOLOv8 and tested with 351 images of 6 bird species. The model increased the mAp50 from 0,954 to 0,984 and the mAP50-95 from 0,633 to 0,781.