Unmanned aerial vehicle swarms (UAVSs) can carry out numerous tasks such as detection and mapping when outfitted with machine learning (ML) models. However, due to the flying height and mobility of UAVs, it is very difficult to ensure a continuous and stable connection between ground base stations and UAVs, as a result of which distributed machine learning approaches, such as federated learning (FL), perform better than centralized machine learning approaches in some circumstances when utilized by UAVs. However, in practice, functions that UAVs must perform often, such as emergency obstacle avoidance, require a high sensitivity to latency. This work attempts to provide a comprehensive analysis of energy consumption and latency sensitivity of FL in UAVs and present a set of solutions based on an efficient asynchronous federated learning mechanism for edge network computing (EAFLM) combined with ant colony optimization (ACO) for the cases where UAVs execute such latency-sensitive jobs. Specifically, UAVs participating in each round of communication are screened, and only the UAVs that meet the conditions will participate in the regular round of communication so as to compress the communication times. At the same time, the transmit power and CPU frequency of the UAV are adjusted to obtain the shortest time of an individual iteration round. This method is verified using the MNIST dataset and numerical results are provided to support the usefulness of our proposed method. It greatly reduces the communication times between UAVs with a relatively low influence on accuracy and optimizes the allocation of UAVs’ communication resources.

The realization of multi-modal image fusion requires sufficient cross domain data. Translating the visible images is an effective method to obtain thermal-visible paired images from the visible image domain to the thermal image domain. The current translation methods have some disadvantages, such as unreasonable distribution of thermal radiation intensity, blurred edges, spatial distortion and feature loss. So they are not friendly to downstream tasks. Based on the generation and reconstruction strategy of CycleGAN, we propose an image to image translation network guided by heatmap which is called HGGAN. We use the heatmap of object that detected by network to encode the heatmap code, and combine the image edge code to improve the image generation performance. We test the image standard of the generated image, and use the object detection network to verify.