Aircraft remote sensing image target detection based on shallow feature enhancement and k-means clustering algorithm

Abstract

Target detection is one of the basic tasks in the field of computer vision. The target detection of aircraft remote sensing images has important research value and application value. Aiming at the problem that the accuracy and real-time performance of current aircraft remote sensing image target detection algorithms can not balance, this paper presents a target detection algorithm aircraft remote sensing image based on the Single Shot MultiBox Detector(SSD). For the problem of poor accuracy of the SSD algorithm for aircraft remote sensing target detection, especially for small aircraft targets, the paper designs a module for shallow feature enhancement. In the end, the shallow network obtains feature information with rich structure and wide perception field, which improves the detection accuracy of the network for small remote sensing targets on low-level feature maps. In order to enable the neural network to learn more effective information, first use the k-means clustering algorithm clustering algorithm to obtain suitable anchor size for the aircraft target, then use the Non-maximum suppression(Soft-NMS) in the post-processing part of the SSD algorithm and focus classification loss function. Soft-NMS can reduce the missed detection of aircraft targets, and the focus classification loss can solve the problem of imbalance between positive and negative samples to a certain extent. Related experiments on the aircraft remote sensing dataset, average accuracy reaches 91.88% and frame per second is 39.3. The results show that the improved SSD algorithm can balance detection accuracy and real-time performance at the same time.