A Multimodal Explainable Deep Learning Approach for E-Commerce Recommendation and Behaviour Prediction

The detection of diminutive targets in UAV aerial images has many important roles and application values. However, due to the small targets and complex terrain occlusions, accurately detecting these targets in UAV images remains challenging. This paper explores effective architectural designs and improve the benchmark model YOLOv8n through three strategies to ensure the increase of detection accuracy without additional computational and parameter overhead. Firstly, an improved C2f module is introduced, which uses partial convolution in place of traditional convolution to reduce model parameters and improve processing speed, and combines Triplet Attention improve the model's capability in understanding complex scenes. Secondly, a specialized header designed for 160x160 small target detection has been integrated into the Neck network. Additionally, ASFF feature fusion technology is employed to dynamically optimize the feature hierarchy using adaptive weights, enhancing the utilization of detailed information. Finally, the SPPF module is optimized by introducing a 13x13 average pooling layer and LSKA large separable kernel attention mechanism, significantly enhancing feature expression of small targets. Compared with YOLOV8n, the improved algorithm on the VisDrone2019 dataset improves mAP@0.5 and mAP@0.5:0.95 by 2.67% and 2.25%, respectively, while the number of parameters decreases by 11.7%.Experimental results confirm that these optimizations substantially improve detection accuracy and speed for small and medium-sized objects in UAV imagery, validating the effectiveness and practicality about these proposed improvements.