Insect pest detection plays a crucial role in agricultural production for accurate and early pest control, thus significantly reducing crop damage and increasing yields. However, currently the small size and multi-scale characteristics of insect pests pose significant challenges for accurate object detection using computer vision technology. To address this issue, we propose a novel framework called DAMI-YOLOv8l to detect pest in images collected by a light-trapping device. The DAMI-YOLOv8l model integrates three key innovations: the Depth-wise Multi-Scale Convolution (DMC) module, the Attentional Scale Sequence Fusion with a P2 detection layer (ASFP2) neck structure, and a novel bounding box regression loss function named Minimum Point Distance inner Intersection over Union (MPDinner-IoU). The DMC module improves multi-scale feature extraction to enable the effective capture and merging of features across different detection scales while reducing network parameters. The ASF-P2 neck structure enhances the fusion of multi-scale features while preserving critical local information related to small-scale features. Additionally, the MPDinner-IoU loss function optimizes feature measurement for small insect pest datasets by introducing geometric correction capabilities. By leveraging these innovations, the results demonstrate that the proposed framework improves many metrics, such as mAP50 from 74.5 % to 78.2 %, mAP50:95 from 52.5 % to 57.3 %, and FPS from 109.89 to 121.12, compared with those of YOLOv8l model on the proposed LP24 dataset. Furthermore, we validate its robustness on two other public datasets related to small objects, Pest24 and VisDrone2019.