Toward Reliable Mango Leaf Disease Recognition with Lightweight Convolutional Networks
Year: 2026
Type: conference-paper
Abstract: Mango leaf diseases can severely reduce orchard productivity and fruit quality, but field diagnosis still relies largely on manual inspection. Manual assessment becomes difficult to scale when symptoms are subtle or images contain illumination changes and background clutter. Reliable automated recognition therefore requires evaluation protocols that reflect generalization to unseen leaves rather than benefiting from near-duplicate leakage. To address this issue, we introduce a leakage-safe pipeline that groups visually similar images using perceptual hashing before stratified group-level partitioning. Using the resulting partitions, we fine-tune lightweight ImageNet-pretrained CNN backbones for eight-class for the classification of mango leaf disease. On a dataset of 4000 images, perceptual-hash grouping produces 3560 groups and a disjoint 2407/803/790 train/validation/test split. Under a fixed training protocol, SqueezeNet1-1 achieves 99.63% accuracy with a macro-F1 score of 0.992, while ShuffleNetV2-x0.5 attains 99.40% accuracy with a macro-F1 score of 0.991. Both checkpoints remain below 3 MB; SqueezeNet1-1 and ShuffleNetV2-x0.5 require 0.18 ms and 0.45 ms per image, respectively, on an NVIDIA Tesla P100 GPU. Overall, the results show that duplicateaware partitioning supports credible performance estimation while preserving the compactness and computational efficiency of lightweight CNNs under the evaluated setting.
Institutions: GITAM University, Jain University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7203645702