Improved Small Target Detection Algorithm Based on SSD
Yue Hu, Quan Zhang · 2024
In response to the issues of small targets being easily occluded and having a high false negative rate in complex indoor detection environments when using the SSD algorithm, this paper introduces an enhanced variant of the SSD algorithm (M3-ECA-SSD). MobileNetV3 is used as the backbone network of the SSD algorithm; the attention mechanism in bneck is replaced with the ECA attention mechanism, and the extracted high and low level feature maps are fused using the FPN structure. The M3-ECA-SSD algorithm is applied to the ADE20k dataset, and the experimental results show that the algorithm in this paper significantly improves the detection ability of the SSD algorithm for small targets, with a mAP value of $82.7 \%$ and FPS of $\mathbf{5 7 . 7 8}$.