Aiming at the issues of high onboard core computing requirements and low computational efficiency in UAV multi-target recognition neural networks, this paper proposes a lightweight neural network designed to reduce the algorithm's reliance on hardware memory and to meet the demands of unmanned equipment for synchronized, lightweight, and highly efficient neural networks. A multi-channel splitting convolutional computing strategy is introduced to decrease the amount of serial computation. A dual allocation strategy is employed to enhance the capability of dense target selection and to reduce the inference process. Angular loss is utilized to address the misalignment between prediction frames, thereby improving the model's inference accuracy and convergence rate. Experimental results demonstrate that the proposed algorithm achieves a recognition accuracy of 94.3%. Compared with the mainstream models YOLOV5n and YOLOV8s on HIT-UAV data set, the experiment shows that the recognition rate of small targets is 2.8% and 0.9% higher. The recognition rate of medium target is 2.2% and 1.3% higher. Enabling UAVs to possess efficient and precise end-to-ground target recognition capabilities even with limited computational resources.