Research on efficient pest identification system for edge computing terminals based on Transformer-ConvLSTM

• Published in: Discover Computing Vol. 28; no. 1; p. 92
• Main Authors: Chu, Shiwei, Bao, Wenxia
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 25.05.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial neural networks; Complexity; Computer Science; Constraints; Consumption; Data Mining and Knowledge Discovery; Data Structures and Information Theory; Edge computing; Identification systems; Inference; Information Storage and Retrieval; Natural Language Processing (NLP); Neural networks; Optimization; Pattern Recognition; Performance evaluation; Transformer-ConvLSTM; Large model compression; Adaptive deployment; Lightweight; Pest recognition; Edge computing
• ISSN: 2948-2992; 1386-4564; 2948-2992; 1573-7659
• DOI: 10.1007/s10791-025-09612-3

This study proposes a domestic edge computing terminal pest identification system based on large model compression and lightweight technology, integrating new algorithms based on Transformer and ConvLSTM, and optimizing its performance in resource-constrained environments through adaptive deployment strategies. By evaluating the performance of four different models (Transformer ConvLSTM, CNN, LSTM and Transformer) in terms of recognition accuracy, inference speed, computing resource consumption and adaptive deployment effect, this study reveals the advantages and disadvantages of each model in edge computing tasks. The model based on Transformer ConvLSTM performed best in recognition accuracy, reaching 94.5%, significantly better than other models; while the CNN model was the most efficient in inference speed, reaching 52 frames per second (FPS). The adaptive deployment strategy of the Transformer ConvLSTM model, the Transformer ConvLSTM model can dynamically adjust the model complexity to effectively cope with higher computing resource requirements, thereby improving inference speed and reducing resource consumption. On resource-constrained devices, the optimized Transformer ConvLSTM model can maintain high recognition accuracy and significantly reduce latency. Although the compression strategy slightly affects the accuracy, it greatly improves the inference speed, especially in the CNN model. Highlights Innovative hybrid model design: It is proposed to combine Transformer with ConvLSTM to design an efficient pest recognition model that can capture global information and temporal features at the same time. Transformer handles long-distance dependencies, while ConvLSTM better handles spatiotemporal data. This fusion enhances the adaptability and robustness of the model in dynamic agricultural environments. Lightweight optimization technology: Through optimization through model quantization, pruning, and knowledge distillation, the model can effectively reduce computing and storage requirements while maintaining high accuracy, making it adaptable to resource-constrained edge computing devices. Adaptive deployment strategy: It proposes an adaptive deployment strategy that dynamically adjusts the model complexity and computing accuracy, optimizes the model deployment in real time according to the hardware resources and network conditions of the edge device, and ensures that the system can achieve optimal performance in different environments.