An Improved Method of Object Detection Based on Chip

Abstract

In spite of methods for object detection based on convolutional neural networks, there's a problem that the information of objects missing in the convolutional progress with an immeasurable proportion. The reason is that while the network downsample in order to further obtain the abstract features, a certain pixel point in the feature map corresponding to more original image area, so there're less content that can be referred to. To handle this problem, an improved object detection method based on YOLOv3 is demonstrated. Our approach is composed of three steps, initial detector, adaptive chip generator, secondary detector. Firstly, figuring out which chips are worth detecting in the image. Secondly, screening the best associations for reduce the number of duplicate detections from these chips. Finally, detection progress will run on each chip and summarize the output. Benefit from it, this method achieves a significant performance especially in medium and large size objects.