In the process of drone aerial photography, the drone monitors and shoots from a top-down perspective, with a relatively single feature extraction angle. Moreover, the drone is far away from the ground, and the target to be detected appears relatively small. When it is obstructed by the background, it is prone to false detection during detection. Therefore, design a lightweight unmanned aerial vehicle aerial small target intelligent detection system based on deep learning. In terms of hardware, a drone aerial photography device has been designed, including a flight control module, a servo gimbal video system, a ground remote control, and a handheld mobile terminal device, to ensure the stability and efficiency of image acquisition; And adopt a 2.4GHz transmitter with high receiving sensitivity and low power consumption to design image transmission equipment, ensuring the effective transmission range and stability of images. In terms of software, a six direction gradient decision method is used to establish a background suppression model, and a weighting function is introduced to process unmanned aerial vehicle aerial images, eliminate the influence of complex backgrounds on intelligent detection of small targets, and improve detection accuracy. Divide the image after background suppression processing, and select several sub regions belonging to the region of interest through judgment analysis. Then, perform stitching processing to obtain the recognition result of the region of interest in the image. Combining UavdNet network and CASA attention mechanism, further analyze the region of interest image and generate multi-scale feature images. By using the fused Stem module and ShuffleNet V2 unit to reconstruct the YOLOv5 backbone network, a lightweight intelligent detection model based on deep learning is established. The generated multi-scale feature images are input into the model for learning, and small object detection results can be obtained. The test results indicate that the AP value of the intelligent detection output by the system is greater than 0.9, which can achieve accurate description of small targets.