Small object detection algorithm based on improved YOLOv8

In the reseach field of targets detection, small targets detection is an important difficulty. Existing target detection algorithms have good performance in detecting medium and large targets, while detection accuracy in smaller target detection is low. To enhance the accuracy of small target detection, we optimized YOLOv8 and proposed a new model based on small targets. First, the shallow feature map with 4 downsampling is used instead of the deep feature map with 32 downsampling, so that the feature map contains more feature information. Secondly, Involution module is introduced to make up for the defect that convolutions can only extract local information, so that the network can obtain rich context information. Finally, an asymptotic feature pyramid is used at the neck of the network for feature fusion to deal with the issue of feature information loss and multi-target conflict caused by feature in the process of information interaction. EMA attention mechanism is added before the detection head to make the network pay more attention to small objects and solve the problem of background information interference, thus easing the question of missing information and false detection. The proposed improvement improves detecting the precision of small objects detection task. We know from the experiment that our model has fewer parameters, and the average accuracy is 4% higher than YOLOv8 model.