Research on Lightweight Deep Learning Real-Time Image Recognition Methods for Drone Aerial Photography

Abstract

The problem is that there is little time and the resources on board the drone are very limited. The attributes of aerial images are not uniform; that is to say, there is an inherent trade-off between model accuracy and inference speed, there are differences in perspective and considerable variations in scale, and there is a high density of small targets. Based on the problems mentioned above, some lightweight deep learning methods for real-time image recognition on drones will be introduced in this paper. At the level of the backbone network, in order to improve the feature learning ability and reduce computational cost, we have modified the structure of depthwise separable convolutions and introduced feature reuse and dimensionality reduction. A hierarchical feature adaptation fusion strategy is introduced in the detector stage of this paper to expand the range and increase the expressiveness of features, and also to solve the problem of failing to detect small targets. Channel pruning is performed at the time of deployment, and at the same time, fixed-point quantization and operator kernel tuning are carried out to increase both the model compression ratio and the inference speed; thus, this paper proposes a relatively small, high-accuracy and fast-running solution for real-time drone aerial recognition.