Object Detection Algorithm based on Dense Connection
PAN Zhihao, CHEN Ying · 2019
The way that information propagates in neural networks is of great importance. In this paper, we propose a connectivity pattern: dense connection, aiming to solve object detection algorithm YOLO-Tiny with less convolutional layers, low feature utilization rate, low precision and poor detection of small objects. We integrate dense connection into YOLO-Tiny, increasing its convolutional layers and improving the feature extraction network. Improved network extracts feature maps and fuses the feature maps by using the Dense Block module. Detection network completes the classification and location at different scales with different anchor boxes. We tested improved network on the Pascal VOC dataset. The experimental results show that our network has improved accuracy by 15% compared with the original algorithm. Although the detection speed has increased, it can still meet the requirements of real-time detection. Compared with the YOLO-Tiny model, our model size only increases by 9.8. MB, compared to the YOLO model, the model size is about 1/5 of the original.