A convolutional neural network-based impairment detection method. An input image can be received, and convolutional multi-feature mappings of different scales are generated; the generated convolutional feature mappings are processed by means of a dual/multi-region proposal network, a dual/multi-impairment proposal is generated for each candidate impairment in the image, and a dual/multi-region proposal bounding box is created; the dual/multi-region proposal bounding box is projected back to the feature mappings of respective convolutional layers to obtain a group of dual/multi-regions of interest; the dual/multi-regions of interest are compared, and a confidence score is created to indicate the likelihood that a desired impairment is detected in the bounding box, so that the desired impairment can be detected only by one step. The beneficial effects of the present application are: less time is spent, the precision and the recall rate are high, the size of a data set is increased, and the convolutional layer can increase the speed of a model and improve the precision to an average precision mean of up to 98% to 99%.
展开▼
