An Evaluation on Effectiveness of Deep Learning in Detecting Small Object Within a Large Image

Multiple Deep Learning (DL) algorithms have been developed recently and are shown to be achieving very high accuracy in object detection. However, challenges have been reported in detecting small objects within a large image (e.g. > 2000 by 2000 in resolution). Various methods have been proposed using different detection algorithms in order to detect small objects. However, these approaches require high computational resources and are not suitable for edge computing devices that are used for practical applications such as pedestrian traffic light detection. We explored two different methods of detection to evaluate which method is best at detecting small objects. The first method is a two—part procedure with the first step being image processing and the second step, a R-CNN based detection using Edge Boxes algorithm for the extraction of region proposals. The second method is solely Faster R-CNN Object Detection with Instance Segmentation, termed as Mask R-CNN. A total of 4000 streets images of Singapore with pedestrian traffic lights were used as training data. The dimensions of the images range from 1200 by 900 to 4000 by 3000. The small object to be detected is the green or red man within pedestrian traffic lights. We evaluated these methods based on training time required, detection time, accuracy as well as suitability for deployment in edge computing devices. From the results, it is shown that the HSV + R-CNN approach is preferred as it achieves an accuracy of 95.5% and can be deployed in edge devices.