Intelligent image identification model of crop diseases and pests based on multimodal deep feature fusion
With the development of smart agriculture, it has become a research hotspot to realize rapid and accurate identification of crop diseases and insect pests by using deep learning technology. In this study, a multimodal deep feature Fusion (MMDFF) optimization model for intelligent identification of crop diseases and insect pests is proposed. Based on the classic AlexNet, the study introduces transfer learning technology and makes full use of the prior knowledge obtained from the pre-training of large-scale general datasets such as ImageNet, which effectively alleviated the problem of insufficient labeled data in the agricultural field. On this basis, a multi-task learning framework is constructed, which combines pest species identification and severity assessment to optimize, and improves the parameter efficiency and feature sharing ability of the model. To further enhance feature representation, it proposes a MMDFF mechanism. The mechanism jointly employs visual features from multiple convolutional neural network layers and texture features extracted by local binary patterns. A cross-modal channel-attention module dynamically calibrates the importance of each feature channel. Consequently, the network focuses on the region’s most relevant to disease identification. To meet the requirements of edge computing, this study also implements a system lightweight scheme including structured pruning and mixed precision quantization. The experimental results on the published Plant Village dataset show that the proposed model achieves an identification accuracy of 95.12% and a mean square error of 0.069 in the test set, and its comprehensive performance is better than the original comparative models such as AlexNet, ResNet-50 and Mobilene ETV3. The ablation experiment further verifies the effectiveness and synergistic effect of each module. After lightweight, the parameters of the model are reduced to 9.3 M, and the single-frame inference delay is 32.7 ms on embedded equipment, which provides a feasible technical path for real-time and accurate pest monitoring in the field.