Chip on carrier (COC) is an important component of transmitter optical subassembly and is widely used in the field of optical communication. With the progress of chip manufacturing process, COC is developing towards miniaturization and high density, and the types of defects become more complex and diverse. Optical inspection technology based on traditional image processing methods can no longer meet the requirements of COC multi-category defect detection. In this paper, YOLOV3 is introduced into the detection of typical defects of COC, such as collapse, positioning column damage, and waveguide stain. Aiming at the problem that the waveguide stain defect target is small and the scale changes greatly among different types of defects, the original YOLOV3 feature extraction network is improved, and the 4 detection scales are designed taking into account the multi-scale characteristics of the target, and the multi-scale detection is improved by enhancing feature fusion. The K-means method is used to perform cluster analysis on the data set, and select the optimized initial prior frame. Experimental results show that the accuracy of YOLOV3-COC, a COC defect detection method based on improved YOLOV3, is 97.4% for the detection of 3 types of defects: COC chipping, broken positioning pillars, and waveguide stain.