Aiming at the problems of small size and complex background of infrared targets detected at long distances in airborne aerial scenes, the average recognition rate is low, and the real-time performance of the airborne infrared target recognition algorithm is high. In this paper, an improved EfficientDet infrared small target recognition algorithm was proposed. First, an Unsharp Masking method suitable for small infrared target data enhancement is used to enhance the target edge and detail information; Secondly, GhostNet was used as the backbone network of the model, which greatly reduces the amount of parameters of the model and improved the inference speed of the model; Finally, the Coordinate Attention module was used. It was applied to the feature layer extracted by the feature extraction network, which increased the feature expression ability of the network and prompted the network to accurately capture the position information of the small infrared target. The experimental results show that the average accuracy of the optimized algorithm in pure background, complex background, small-scale target and multi-target application scenarios is improved by 2.24%, and the FPS reaches 46.56. Small target recognition provides the basis for edge computing.
There are deviations in the images of the same target at different angles, so the multi-angle infrared images contain additional classification and identification feature information. Therefore, the multi-angle image acquisition system can effectively solve the problems such as occlusion and small target size that cannot be recognized with high precision. Combined with the multi-angle acquisition system, a multi-angle infrared vehicle target recognition algorithm based on Light-Head R-CNN was proposed. Firstly, inputted the multi-angle infrared images of the same target into the ConvNeXt backbone network for feature extraction; Secondly, in order to reduce the amount of model parameters and realize the lightweight of the network, the traditional convolution in the backbone network was replaced with Ghost Module, which reduced the amount of FLOPs and parameters by about 50%; At the same time, according to the characteristics of infrared vehicle targets, the Light-Head R-CNN target recognition algorithm was improved, and an ultra-lightweight ECA module without dimensionality reduction local cross-channel interaction was added to improve the performance of the network for infrared vehicle target recognition; Finally, the Dempster synthesis rule was used to perform data fusion on the recognition accuracy of images from different angles predicted by the network to obtain the final recognition accuracy. It had been verified that compared with single-angle images, when the number of input angles was 2, the recognition accuracy of the algorithm in the constructed data set was improved by 3.8%; The recognition accuracy reached 90.1%, it was optimal, while the input infrared images were distributed at a certain angle and the number was 5, at the same time the speed achieved 44fps. The results fully demonstrated the feasibility of the proposed algorithm in improving the recognition accuracy, and provided a theoretical and experimental basis for enhancing the performance of the target recognition algorithm.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.