MC-YOLOv5: A small target detection model based on YOLOv5
Abstract: Small object detection is one of the most challenging tasks in object detection. The small size of objects, low resolution, and occlusion introduce difficulties in accurately detecting them. To address these challenges, we propose an improved small object detection model called MC-YOLOv5, based on YOLOv5. In MC-YOLOv5, we replace the backbone network with MobileNetV3 to enable the model to perform efficiently on mobile devices and better adapt to small object detection tasks. Furthermore, we enhance the feature extraction capability of the model by incorporating the CBAM (Convolutional Block Attention Mod-ule) attention mechanism into the detection head. In addition, we improve the data augmentation techniques by extending the traditional Mosaic data augmentation from combining 4 images to 9 images, thereby reducing the object size. We also introduce a synthetic fog algorithm to add masks to the images, which reduces the resolution and simulates noise. Experimental results on the COC02017 dataset demonstrate that our MC-YOLOv5 model performs better in small object detection. Compared to the traditional YOLOv5, our model shows an approximate 7% improvement in accuracy and significant enhancements in metrics such as mean average precision (mAP).
Loading