Meta-learning few shot object detection algorithm based on channel and spatial attention mechanisms

The two-branch based meta-learning few shot detection algorithm will first use the shared backbone network to get the support features and query features, use the query features for bounding box generation and get the bounding box features, then use the aggregation algorithm to aggregate the bounding box features with the support features, and then use the aggregated features to completion classification and regression tasks. Since the simple channel multiplication aggregation algorithm misses valuable high-level semantic information such as channel and space, it performs poorly in the detection of novel classes of targets. To address this problem, a meta-learning few shot object detection algorithm based on channel and spatial attention mechanisms is proposed. The algorithm proposes a convolutional attention mechanism module with residual block, which includes a channel attention mechanism and a spatial attention mechanism. The support features processed using this module better characterize the semantic information that there are individual categories on channel and spatially. Aggregation of this information with bounding box features can be used to better accomplish classification and detection tasks. In addition, a decoupling module for classification and regression is proposed, which decouples the classification and regression tasks to improve the performance of the algorithm. In this module, different aggregation algorithms are used for classification and regression tasks respectively. To verify the effectiveness of the algorithm, experiments are conducted on the public dataset PASCAL VOC. The experimental results show that the performance of the algorithm gets a performance improvement over the baseline model.