An Improved SSD Seal Detection Algorithm Based on Channel-Spatial Attention

In response to the challenges in seal detection due to small seal targets, faint imprints, and text occlusions, which result in low detection accuracy and difficulties in detection, an improved SSD (Single Shot multi-boxes Detectors) seal detection algorithm based on channel-spatial attention is proposed, achieving fast and accurate detection of document seal images. Firstly, the Channel-Spatial Attention Module (CSAM) is introduced to separately obtain channel attention and spatial attention features from multi-scale features. The Attention Fusion Layer (AFL) is used to integrate context information of features from different scales; combined with the feature fusion layer, it enhances the expression of features to be detected. Secondly, Shuffle Net V2 is introduced as the backbone network to replace the VGG network, effectively improving the detection speed of the model. Additionally, this paper constructs a seal detection dataset specifically for seal detection and proposes a random occlusion seal image augmentation strategy. Experiments show that compared to the original algorithms, the improved algorithm increased detection accuracy by 5.2% points and improved the inference speed by 14 FPS, which essentially meets the practical needs of the project.