Study Of UAV Target Recognition Based On Deep Learning

Modern warfare is the basic driving force to promote the development of UAV.UAV operates through the perception of scene information,thus performing boring or dangerous tasks with less human intervention.Therefore,UAV target recognition is one of the most important development directions in UAV system.In this paper,the target recognition of fixed-wing aircraft in UAV aerial images was taken as an application scenario.Based on the classical research results in the field of computer vision,the deep-learning-based target recognition algorithm was optimized and innovated.This paper mainly works as follows:1.Preprocessing Images.The brightness information of color image was extracted and the gray scale image was generated.According to the neighborhood information,the BMS saliency region of the image was enhanced by non-linear self-adapting to improve the quality of over-exposed and underexposed images.Thus,the influence of unsatisfactory illumination conditions on the results in target recognition technology was reduced.2.Based on the framework of Faster R-CNN,a target recognition method with VGG16 network and region proposal was studied.During the application scenario,three factors were improved—the activation function of convolutional layers to extract features,the generation and merging strategy of preferred windows of region proposals and the loss function of the multi-vision.Furthermore,the relevant parameters of VGG16 network,RPN and softmax classifier were also adjusted.To realize the training of deep network,the NWPU VHR-10 dataset was enhanced.Finally,the network model was validated on UCAS-AOD dataset.The recognition effect proved the reliability and superiority of the proposed network model.3.Combining with the real-time super-resolution reconstruction sub-network,a target recognition neural network for small-scale fixed-wing aircraft was designed.Selective pruning of superresolution reconstruction sub-networks can accelerate the learning process and avoid over-fitting.Experiments shew that the optimized network can improve the recognition accuracy and recall rate.To a certain extent,it overcame the difficulties of small target recognition in UAV target recognition.