Aiming at the low efficiency in the detection of inferior cocoons during the cocoons processing due to manual visual inspection, a method based on machine vision was adopted to detect inferior cocoons. Firstly, according to the depth of field of image acquisition system, appropriate shooting distance for the line scan camera was selected, and further the parameters of the image acquisition system were configured based on the sampling frequency. Secondly, the inferior cocoons detection data set was constructed based on the area array images obtained by synthesizing the linear array images. Finally, a inferior cocoons real time detection model (inferior cocoons net, ICNet) was designed based on YOLO v4 target detection model. The model used the K-means algorithm to perform cluster analysis on the data set of inferior cocoons to preset the candidate anchor parameters and improve the model accuracy. By adopting the method of model depth manipulation, the model was compressed to achieve lightweight and fast detection speed. In addition, the lightweight convolution module was designed for a lightweight feature extraction network to further improve the speed of the model. Compared with the original YOLO v4 basic model, experimental results showed that the mean average precision of ICNet inferior cocoons real time detection model was improved by 1.87 percentage points to 95.55%, the storage space occupied by the model weight was compressed by 40.82% to 145.00MB, and the average detection speed was improved by 91.65% to 49.37 frames/s.