Durian (Durio zibethinus) is a high-value fruit crop in Viet Nam, with rapidly increasing production, which underscores the need for accurate leaf condition recognition in real orchards. Field images are challenging because symptoms vary in scale and are captured under uncontrolled illumination and complex backgrounds, while many deep models improve accuracy at the cost of higher computation. This study proposes an efficient YOLOv8-based classifier for durian leaf images. The method removes the deepest downsampling stage to preserve higher-resolution features and introduces PartialTriDW, a lightweight tri-branch depthwise block with adaptive softmax fusion and low-cost channel mixing, to enhance multi-scale texture representation in the early and mid-backbone stages. Experiments on a five-class dataset of 4,437 field images, trained from scratch, show that the YOLOv8 baseline achieves 0.918 accuracy, the CutP5 variant achieves 0.927 accuracy, and the proposed model reaches 0.949 accuracy with 0.56 million parameters and 3.3 GFLOPs, while maintaining comparable latency. The results indicate that preserving spatial detail and strengthening multi-scale texture modeling improve practical durian leaf classification with controlled computational cost.