Enterprise AI Analysis: Computer Vision & Remote Sensing

The Efficient Multi-scale Cross-Attention (EMCA) mechanism is integrated into the YOLOv8 backbone to address the limitations in capturing fine-grained visual features of infected foliage. PWD often manifests with subtle color variations and indistinct structural features, easily confounded by complex backgrounds. EMCA enhances the model's sensitivity to these pathological cues by facilitating cross-channel and cross-scale feature interaction and fusion.

This lightweight attention module (see Figure 4 in the paper) performs parallel convolutional operations at multiple scales to capture varied receptive fields, improving detection across different resolutions. By dynamically adjusting response weights of critical feature channels, EMCA effectively highlights disease-affected regions while suppressing redundant background noise. The ablation study showed EMCA improved AP@0.5 by 2.8% points and AP@0.5:0.95 by 4.5% points.

The conventional detection head of YOLOv8 is replaced by the Self-Ensemble Attention Module (SEAM) to improve robustness in complex forest environments, particularly for identifying occluded and overlapping diseased crowns. SEAM (see Figure 5 in the paper) employs a dual-branch architecture that jointly models spatial details and high-level semantic features, enhancing the model's discriminative capability.

One branch uses dilated convolutions for semantic enhancement, capturing long-range dependencies and macro-level lesion characteristics. The other branch maintains original spatial details via shallow convolutions, focusing on fine-grained details like edges and textures. Outputs from both branches are fused with channel and spatial attention mechanisms. This not only enhances feature extraction but also contributes to model compression, reducing parameters by 0.48 MB and improving AP@0.5 by 0.4% points and AP@0.5:0.95 by 0.7% points after EMCA integration.

To overcome challenges in accurately localizing difficult samples and improve discrimination of complex samples, the traditional CIoU loss function is replaced with the novel Focaler-IoU loss function. This improved loss function (see Figure 7 in the paper for comparisons) incorporates dynamic loss weighting, monotonicity constraints, and advanced IoU optimization.

Focaler-IoU increases the model's focus on hard-to-localize samples, modulating penalty strength according to sample difficulty. This refines localization accuracy and enhances robustness, particularly for challenging cases involving occluded or blurred target boundaries. Compared to the previous stage (EMCA+SEAM), Focaler-IoU improved AP@0.5 by 0.8% points and AP@0.5:0.95 by 2.1% points.