Multi-dimensional feature fusion network design and performance optimisation for small target detection

Due to the long distance of image acquisition, high imaging resolution, complex feature background, shooting angle, etc. The result is that there are few features available for small targets and they are easily interfered by background noise, which poses a challenge to the detection of small targets. To address the above problems, this paper proposes a target detection network (Convolution-based Small Target Detection Network, CSTDNet) with enhanced feature information, which integrates a multi-dimensional information fusion strategy for small target features. An all-round efficient feature fusion mudule (AeFusion) is introduced, which emphasises the fusion of multi-dimensional feature information, enhances the model's ability to focus on key information and suppress redundant information, and strengthens the ability to characterise local features and details, improving the effectiveness of the information and computational efficiency. In order to further enhance the location-awareness capability in cross-layer interaction, this paper introduces a novel decoupling head (Self-aware task decomposition for fine-grained feature sharing, STFS), which improves the accuracy of the small-target classification and localisation tasks through efficient detail sharing and task auto-alignment functions. And localisation tasks through efficient detail sharing and task auto-alignment. This study evaluates the effectiveness of the algorithm on five different scenarios containing small target datasets. Experimental results show that CSTDNet achieved improvements of 6.6 %, 5.8 %, 5.8 %, 5.5 %, and 5.6 % over the baseline model in terms of the mean average precision ([email protected]) metric on the Visdrone 2019, BDD100K, WiderPerson, SODA10M, and AppleDatas datasets, respectively, demonstrating stronger detection performance.