Tiny Object Detection with Lightweight YOLOv5 Based on Attention Mechanism

In object detection tasks, Frontier small object detection algorithms basically suffer from low accuracy in detecting small objects. Given this situation, we propose the lightweight YOLOv5 small object detection algorithm it is an attention mechanism based algorithm. Using the depth-separable convolution criterion of the original model to realize the lightweight of the model, we modify the neck of YOLOv5 and combine FPN and PANet to fuse different semantic information in four feature maps, thus improving the quality of feature extraction and the performance of small object detection. We modified the main feature extraction network of YOLOv5. Based on the original output of three feature maps, Incorporate new feature images so as to achieve feature enhancement of the initial image. We replace the original ReLU activation with Swish activation using an activation function that is more suitable for retaining the target function. To enrich the target detection environment, we applied the data augmentation method, Here we used a dynamic update of parameters using cosine annealing for training of parameters and ported it to our newly investigated YOLOV5 algorithm. And it was validated effectively and allowed to experiment on the YOLO family of algorithms in the CityPrersons dataset. The results show that our innovative YOLOv5 algorithm can be very effective for the detection of small objects.