Research On Lightweight Aerial Object Detection Algorithm Based On Deep Learning

Object detection is one of the computer vision fields that have been widely used in recent years.With the increasingly mature technology of unmanned aerial vehicles,its development has opened up a new field of computer vision.UAVs with high-definition cameras gradually occupy the market of traditional industries with their unique advantages such as flexibility and convenience,and put forward higher requirements for aerial target detection algorithms.Limited by small target scene and chip performance,the traditional target detection algorithm cannot balance the model accuracy,size and speed,which fails to meet the requirements of actual production.In order to solve these problems,we propose corresponding improvements based on the YOLOv3 algorithm,which not only improves the accuracy and speed of the model,but also greatly reduces the volume of the model.Our main work includes accuracy improvement and model compression.On the one hand,we first analyze the mismatch in the receptive field of the YOLOv3 algorithm in the detection of small targets,and design a targeted adjustment plan for the prediction layer structure;then,in view of the insufficient assistance of the low-level features of the FPN structure to the high-level features,we design the neighbor feature interweaves the enhanced structure;finally,for the single feature fusion method,we introduce the PANet structure as reverse feature compensation.On the other hand,we first draw on the channel pruning algorithm to cut a large number of model parameters;then for the unique continuous residual block structure of Darknet-53,we design a residual pruning algorithm;Finally,in order to effectively improve the accuracy of the pruned model,we innovatively introduce a knowledge distillation strategy to finetune the model.After two improvements,we design a comparative experiment on the enhanced YOLOv3.The experimental data shows under the input of 640 × 640,our optimal model can reach a mAP of 31.9%,a speed of 57fps,and a model capacity of only 3.65M.Compared with the enhanced baseline algorithm,the accuracy is 4.7%higher;compared with the pruning algorithms of SlimYOLOv3-SPP3,our optimal model has a greater improvement in speed,accuracy and model size,which proves the effectiveness of our improvements.