Proposal generation method for object detection in infrared image

Abstract In an infrared image, there is a significant difference between the region of the interested object and its surrounding background. Based on this observation, we propose an effective and efficient proposal generation method which uses a Multi-layer and Multi-size Superpixel Segmentation (MMSS) scheme for object detection in the infrared image. The SLIC (Simple Linear Iterative Clustering) algorithm is applied to partition an infrared image into multi-layer and multi-size superpixels. In each layer, only the individual superpixel and the merging of two adjacent superpixels are used to create the candidate pool of object proposals. A superpixel-based center-surround feature is then defined to measure the discrepancy between the region of the proposal and its surrounding background. To evaluate the performance of the MMSS-based method of proposal generation method, we create an Infrared Interested Object Image Dataset (IIOID), in which the infrared images are collected from several benchmarks and the ground-truth of the interested object segmentation is manually labeled. Compared with several state-of-the-art methods of proposal generation on IIOID, the MMSS-based method has overwhelming superiority in detection recall under different Intersection over Union (IoU) thresholds and is convenient for computation. Furthermore, we implement the MMSS-based method as a processing step for pedestrian detection. Experimental results on benchmark infrared pedestrian image dataset show that the detectors with our method of proposal generation method can greatly reduce the number of candidate windows to be detected and also suppress false positives.