An improved SSD based on feature fusion and attention

摘要

SSD (Single Shot Multi box Detector) is one of the most popular object detection detectors with high precision and fast speed. However, SSD's feature pyramid detection method makes it hard to detects small objects. In this paper, we proposed an improved SSD named FFASSD, which includes a lightweight feature fusion module and an efficient channel attention module. FFASSD can especially improve the performance of SSD in small object detection. In the feature fusion module, features from different layers with different scales are concatenated together, the features of shallow layers are replaced by the generated new features to predict the final detection results. In the channel attention block, capturing local cross-channel interaction before the multi-scale feature maps being classified and regression......