PE-SSD: Improved SSD for Small Object Detection

With the flourishing development of deep learning, object detection technology has achieved encouraging results. However, small objects remain a major challenge in the field of object detection due to their lack of texture and contextual information caused by their small size. Despite such fact, we observed that their edges typically exhibit high contrast in relation to the background. Specifically, we innovatively employ a Laplacian image pyramid incorporated with feature enhancement modules rather than the traditional image pyramid, with the aim of more effectively capturing image details and edge information, given that the Laplacian pyramid has the ability to retain such information within its lower levels. Furthermore, we introduce attention mechanisms to optimize the feature fusion process, thus improving the utilization efficiency of multi-scale features. Taking SSD as the baseline, we propose a novel method called PE-SSD, which integrates the advantages of featureized Laplacian pyramid. Experimental results show that our method achieves higher accuracy in detecting small objects compared to traditional SSD. For inputs of size 512 × 512, the extensive experiments on the VisDrone2019 dataset demonstrates that our method achieves an average precision of 15.4%, surpassing traditional SSD networks by 1.2%. Specifically, The mAP_s score reaches 6.6%, which is 0.4% higher than the baseline network, indicating the effectiveness of our method in the field of small object detection.