Cascaded backbone hardware fitting detection model and pruning method

The invention discloses a cascaded backbone hardware fitting detection model and pruning method. The method comprises the steps of: constructing a hardware fitting data set, and obtaining different types of hardware fitting images; designing a composite backbone network based on a multi-fusion ladder cascade structure; introducing an efficient channel attention mechanism to optimize cascade backbone output to enhance effective features; generating a prediction result through three prediction branches of a central point, offset and scale by referring to an anchor-free network thought; and designing a model pruning method based on sparse training to reduce the parameter quantity and the size. According to the method, an anchor-free idea is applied to the field of hardware fitting detection, the precision is improved by designing a cascade backbone network and an efficient channel attention mechanism, a model compression method based on iterative sparse training is also designed, and maximum-proportion compressi