Drones can achieve fast cruising, transmit images back to the mobile devices of management personnel, and use optical zoom lenses to monitor target areas with artificial blind spots. The use of drones for delivering goods can achieve local express delivery, reduce labor costs, and scale up operations. However, existing algorithms for automated drone delivery still have the drawback of blurred image recognition. This article addresses the existing problems of the aforementioned algorithms and introduces twin networks into the field of multi focus fusion of drone images. This article adopts a three-layer stacked small convolution kernel instead of a large convolution kernel. This not only achieves the same perception as the large convolutional kernel, but also adopts a bottleneck like residual block to replace the convolutional layer in the middle of the twin network. Finally, through experimental analysis, it was found that after improving the network training stagnation, the accuracy of the training network reached 98.76%. The method in this chapter achieved the best performance in three indicators: MSE (Mean Square Error), NCC (Normalized Cross Correlation), and SSIM (Structural Similarity Index Measurement), with values of 71.76, 0.8786, and 0.6337, respectively, which are superior to the other commonly used methods. Experimental data shows that the fusion results of the algorithm proposed in this paper have superior visual quality compared to recent algorithms, and the quantitative analysis index values are more prominent than recent algorithms, demonstrating competitive superiority. This has brought new research results to the research of drone delivery technology.