Searching ROI for Object Detection based on CNN

Abstract

Several studies have explored the structural design of CNN to improve the network's performance since a well-designed feature extractor can benefit convolution-based tasks. Although CNNs are able to learn important patterns on raw images, images may contain unpredictable noise that can negatively influence the convolutional stage. Feature extraction cannot always accurately capture the desired features based solely on the input image, but including extra information could improve the result. This paper proposes a fusion input design to generate an additional feature that a CNN can use to provide extra ROI information. Whether a model can utilize the additional information is a determining factor that affects the performance improvement. The proposed method is tested on two public datasets with different structural designs. Overall, the results indicate that additional ROI information can deliver benefits to specific tasks.