Negative Bootstrapping for Weakly Supervised Target Detection in Remote Sensing Images

Abstract

When training a classifier in a traditional weakly supervised learning scheme, negative samples are obtained by randomly sampling. However, it may bring deterioration or fluctuation for the performance of the classifier during the iterative training process. Considering a classifier is inclined to misclassify negative examples which resemble positive ones, comprising these misclassified and informative negatives should be important for enhancing the effectiveness and robustness of the classifier. In this paper, we propose to integrate Negative Bootstrapping scheme into weakly supervised learning framework to achieve effective target detection in remote sensing images. Compared with traditional weakly supervised target detection schemes, this method mainly has three advantages. Firstly, our model training framework converges more stable and faster by selecting the most discriminative training samples. Secondly, on each iteration, we utilize the negative samples which are most easily misclassified to refine target detector, obtaining better performance. Thirdly, we employ a pre-trained convolutional neural network (CNN) model named Caffe to extract high-level features from RSIs, which carry more semantic meanings and hence yield effective image representation. Comprehensive evaluations on a high resolution airplane dataset and comparisons with state-of-the-art weakly supervised target detection approaches demonstrate the effectiveness and robustness of the proposed method.

Weakly Supervised Object Detection for Remote Sensing Images: A Survey

“…The clear improvements obtained by subsequent methods show that the usage of more powerful feature extractors such as CNN allows fully supervised approaches to be outperformed based on handcrafted features (BOV [93] and FDDL [91]). The results are consistent with the literature stating that handcrafted features are not powerful enough to accurately describe objects in RSIs [31,33,34].…”

“…Before the advent of deep learning (DL), most object detectors were based on SVMs. The workflow behind these methods is to start by producing candidate proposals exploiting either a Sliding Window (SW) [31,32,36] or Saliency-based Self-adaptive Segmentation (Sb-SaS) [29,30,33,34] approach. SW generates proposals by sliding, on the entire image, multiple BBs with different scales while Sb-SaS produces saliency maps that measure the uniqueness of each pixel in the image and exploit a multi-threshold segmentation mechanism to produce BBs.…”

“…With the advent of convolutional neural networks (CNNs) [61], both WSOD and RSWSOD methods started to benefit from the powerful feature extraction capabilities of deep architectures. In 2015, Zhou et al [33,34] proposed exploiting transfer learning on a CNN to extract high-level features to feed to an SVM-based detector. The authors further highlight the importance of the process used to select negative instances for training.…”

A survey on object detection in optical remote sensing images

“…Currently, most of geospatial object detection methods, except for few works such as (Han et al, 2015;Tang et al, 2015;Wang et al, 2015;Zhou et al, 2015b), are still dominated by handcrafted features and shallow learning-based features (e.g. SR-based features (Chen et al, 2011b, c;Cheng et al, 2014c;Du and Zhang, 2014;Han et al, 2014;Liu and Shi, 2014;Yokoya and Iwasaki, 2015;Zhang et al, 2014a;Zhang et al, 2015c;Zhang et al, 2015d)).…”

“…Although a few WSL approaches have been developed for natural scene image analysis (Deselaers et al, 2012;Zhu and Shao, 2014), those methods cannot be directly used for geospatial object detection because they are insufficient to address the challenges of RSIs, such as variations in the visual appearance of objects and complex background clutter. As an initial effort, our recent work in (Han et al, 2015;Zhang et al, 2015a;Zhou et al, 2015a;Zhou et al, 2015b) have shown the feasibility of using WSL for geospatial object detection. However, two challenging problems still exist.…”