Inspection image anchor-frame-free target detection method based on neural network architecture search

The invention discloses a routing inspection image anchor-frame-free target detection method based on neural network architecture search. The method comprises the following steps: determining a first-stage anchor-frame-free target detection algorithm FCOS as a bottom-layer target detection method; designing an unmanned inspection image protocol task based on the classic unmanned inspection image data set; constructing a search space; designing search spaces for the FPN structure and the Head structure respectively; evaluating the constructed model according to a preset rule; implementing a search strategy; the search strategy based on reinforcement learning uses the LSTM as an intelligent agent of reinforcement learning, the search space constructed in S300 is used as a state space of a reinforcement learning system, and a classification result which is output by the LSTM in a single step and passes through a Softmax layer is regarded as an action; and performing an experiment on the classic unmanned inspecti