A DirectionalShape-Adaptive Convolutional Kernel Network for Small-Object Detection

In object detection networks based on convolutional neural networks, the performance of small-object detection is affected by multiple factors. One key factor is that, due to the low resolution of small objects, their features are gradually diluted after multiple convolutional operations. Multi-scale pyramid features can alleviate this issue by performing prediction in shallow feature layers, but the improvement is still limited by the fixed sampling structure of convolution: each operation involves a fixed neighborhood, and the more convolutions are applied, the more the object’s own features are diluted by background pixels. This paper proposes a directional shape-adaptive convolution (DSAConv) for image feature extraction. Nine 3×3 kernels with directional displacement and different mask shapes are designed, and the maximum response is selected at each spatial position, providing a response-driven selection that is related to the object shape. The cascaded use of these kernels yields a variable receptive field that depends on the selected kernels, which is beneficial for small-object feature extraction. Integrated into the C3K2 Bottleneck blocks of YOLOv11s, DSAConv improves small-object APS by 2.7 and 3.2 points on COCO and VisDrone, respectively. Extensive experimental results demonstrate the effectiveness of the proposed method.