Efficient Attention-Lightweight Deep Learning Architecture Integration for Plant Pest Recognition

Abstract

Many real-world agricultural applications, such as automatic pest recognition, benefit from lightweight deep learning (DL) architectures due to their reduced computational complexity, enabling deployment on resource-constrained devices. However, this paradigm shift comes at the cost of model performance, significantly limiting its extensive use. Traditional data-centric approaches for improving model performance, such as using large training datasets, are often unsuitable for the agricultural domain due to limited labeled data and high data collection costs. On the other hand, architectural improvements, such as attention mechanisms, have demonstrated the potential to enhance the performance of lightweight DL architectures. However, improper integration can lead to increased complexity and diminished performance. To address this challenge, this study proposes a novel mechanism to systematically determine the optimal integration configuration of popular attention techniques with the MobileNet lightweight DL architecture. The proposed method is evaluated on four variants of two benchmark plant pest datasets (D1_5,869 and D1₅₀₀, D2₁₅₉₉, and D2₅₄₅) and the best integration configurations are reported along with their results. The Bottleneck Attention Module (BAM) attention mechanism, integrated into 12 different layers of MobileNetV2 (BAM12), demonstrated superior performance on D1₅₈₆₉ and D1₅₀₀, and D2₁₅₉₉ and D2₅₄₅, while integrating BAM into eight layers yielded higher accuracy on D2₁₅₉₉. As a result, a comparison with the MobileNet baseline demonstrates that the careful integration of attention mechanisms significantly improves performance.