Leveraging Label-Free and Gradient-Based Pruning for Efficient Neural Networks in Maize Disease Detection

Abstract

Traditional methods for diagnosing crop diseases rely heavily on expert visual assessments, which are often costly, time-consuming, and prone to human error. Deep neural networks, particularly convolutional neural networks (CNNs), have demonstrated strong potential for automating disease detection. However, their deployment remains limited among smallholder farmers due to substantial computational demands and reliance on stable Internet connectivity. To bridge this accessibility gap, lightweight models that operate efficiently on resource-constrained offline devices are essential. Although model pruning is a promising approach for reducing computational overhead and enabling edge deployment, it often results in performance degradation, thereby limiting its practical applicability for high-accuracy crop disease detection. However, to determine the channels contributing little to the model’s discriminative capability but are incorrectly preserved is challenging. This article introduces a novel pruning method based on the gradient analysis of the output difference between pretrained and pruned models concerning noise-perturbed images and the weight parameters. We identify and eliminate less significant weight channels without compromising accuracy by measuring changes in the output of intermediate layers or blocks when small-weight perturbations are introduced. The proposed block weight and image perturbation-induced gradients method is applied to maize disease detection using the faster region-based CNN detector with different backbones. Notably, removing 30.00% of floating point operations and parameters of ConvNeXt results in only 0.62% drop of the Top-1 mean average precision at an intersection-over-union threshold of 0.50 (mAP@50) compared to the vanilla model, particularly when the target model is tiny. It also shows the state-of-the-art performance across various model architectures such as MobileNetV2, ResNet-50, ConvNeXt, and Swin Transformer on standard Canadian Institute for Advanced Research 10-Class Dataset (CIFAR-10) and ImageNet classification benchmark datasets, even when 57.75% of the parameters are pruned. These gains hold consistently across the evaluated one- and two-stage detectors with different backbone architectures, such as CNNs and Transformers.