An Efficient and Light-weight Detector for Wine Bottle Defects

Detecting defects on the wine bottle surface is a challenging task due to these factors: (1) tiny defects on the surface, (2) visually similar defects with design patterns on bottles, (3) reflective (metallic) bottle surface, (4) real-time requirement. In this paper, we propose an efficient and light-weight detector for the defects. To this end, we adopt EfficientNet- B3 [1] as the backbone of the detector and YOLO [2] as the detection head. In addition, we use channel pruning approach [3] to obtain a more compact model. In the experiments, we obtain a score of 0.77 on the validation dataset, similar performance comparing to the two-stage methods[23]. The parameters of our detector are only 5. $8\mathrm{M}$ , which are about one tenth of the YOLO's [2]. The proposed detector achieves a good balance between speed and accuracy.