Attention module distribution in the MobileNetV2 bottleneck blocks for coffee leaf disease classification

The use of deep learning models for detecting plant diseases on mobile devices is still limited by the fact that these devices don't have enough computing power. MobileNetV2 has a lightweight architecture, but using attention mechanisms to improve its feature representation can add extra computational load, especially when used evenly across all layers. To mitigate this limitation, this study introduces a parameter efficient strategy by confining attention modules to 50% of the 17 bottleneck blocks and systematically examines the impact of various spatial distributions on model performance. In particular, four attention mechanisms (CBAM, ECA, SimAM, and Coordinate Attention) are tested with different distribution strategies, such as odd, even, early, middle, edge, and late placements. We used eight different distribution setups and five independent trials to test the stability of performance on a dataset of 1,544 Arabica coffee leaf images divided into three groups: Healthy, Phoma, and Cercospora. The results consistently show that selective and structured attention placement works better than full distribution. This suggests a clear performance trend: Middle > Edge > Odd > End > Start > Even > All > Baseline. The Middle distribution with Coordinate Attention had the best accuracy (0.9174 ± 0.0176), and the Edge distribution with ECA had the best balance between accuracy (0.9161 ± 0.0130) and speed (299.54 GMACs; 4.67 ms inference time). These results indicate that the spatial distribution of the attention module is more critical than simultaneous integration, resulting in enhanced performance at a reduced cost. This study highlights the importance of systematic attention module distribution and lays the groundwork for future research on adaptive methodologies, such as attention-guided pruning and dynamic network architecture search in resource-constrained environments.

Text 126090.pdf - Published Version Restricted to Repository staff only Download (9MB) Official URL or Download Paper: https://linkinghub.elsevier.com/retrieve/pii/S2772...

Actions (login required)

View Item