DeeppestNet: An end-to-end framework for high-performance crop pest recognition

Accurate and efficient crop-pest identification is essential for sustainable agriculture. However, noisy background regions make it difficult to accurately identify crop pests since they obstruct the feature extraction process. Furthermore, pest recognition remains challenging due to limited data and algorithmic complexity. DeeppestNet, a graph pyramid attention-based Bidirectional Long Short-Term Memory (GPA-BiLSTM) model, is proposed as a solution to this problem. In order to improve the recognition process, the Contrast adaptive limited histogram equalization (CLAHE) approach is first used to increase the image quality. Rich feature maps of fine-grained feature regions are intended to be provided by the adaptive pyramid attention module with a cross stage partial (AP-CSP) backbone network. In order to acquire multi-scale spatial features and improve recognition skills through graphical relations, a multi-level pyramid structure is also provided. Graph-based BiLSTM (G-BiLSTM) is used for the final classification, and the Grey Wolf-Salp Swarm Optimization (GW-SSO) technique is used to improve the accuracy. Robust multi-dimensional structural features are extracted with spatial and temporal dependencies when combining G-BiLSTM with a CNN backbone. Additionally, the integration of GW and SSO improves the performance by assuring high precision, fast convergence and balance exploration, exploitation strategies are achieved. The IP-102 dataset is used to assess the proposed pest detection method utilizing evaluation measures like f-measure, recall, precision, and so on. DeeppestNet has achieved 4.6 % higher accuracy than EfficientNet. The experimental outcomes demonstrate that the proposed method performs better than the greatest advanced Deep Learning (DL) algorithms. The proposed method is accurate, efficient, and computationally efficient in comparison to other methods.