This invention discloses a retrieval-enhanced
crop disease detection method based on dynamic memory and human-computer
collaboration, including its implementation
system. The implementation
system comprises a
perception layer, an
edge computing layer, a central service platform layer, and a human-computer
interaction layer. The specific implementation logic of the detection method is as follows: the
perception layer acquires images through field cameras, drones, and other devices; the
edge computing layer performs image reception, preliminary detection, and
feature extraction; the central service platform layer encompasses modules for historical sample retrieval, dynamic sample memory, retrieval-enhanced
inference, real-time correction, and memory update; the human-computer
interaction layer is used for
visualization, positive and
negative sample labeling, and result confirmation. In the implementation process, this method first extracts candidate region features through the
edge computing layer, then retrieves similar historical positive and negative samples from the dynamic memory of the central service platform layer, converts them into memory tokens, and injects them into the
inference process. By enhancing positive samples and suppressing negative samples, the retrieval-enhanced results are output. When the results are questionable, the human-computer
collaboration process is triggered. The
interaction layer receives manual annotations and generates structured memory tokens, enabling real-time correction of
inference and synchronous updating of the memory. This invention enables real-time injection and cyclical reuse of manual corrections, allowing the model to have continuous learning capabilities, significantly improving the accuracy of
crop disease detection and reducing the
false detection rate.