Object Detection for Inventory Stock Counting Using YOLOv5
Isaiah Francis E. Babila, Shawn Anthonie E. Villasor, Jennifer C. Dela Cruz
Abstract
The study has successfully created a program to detect cellphone boxes namely, Cherry Aqua S9 and Cherry Flare S8 in any orientation. Successful detection was made possible by building the datasets where the target objects will capture using a camera in different scenarios. A total of 1,623 images were collected and automatically split by the Roboflow application. In this study, the best background light source to be used is 9Watts. Based on the data gathered were, both target objects had a perfect count of thirty-two (32) and had a 0.96 average accuracy for S8 and a 0.95 average accuracy for S9. For the 7W lighting source, the S9’s did not detect the side view 180° orientation; for the accuracy test, a 0.95 average accuracy for S8 and a 0.90 average accuracy for S9. Lastly, using 5W, eight (8) misdetections in the system, having a total count of only Twenty-four (24) for S8 and forty (40) for S9, for the accuracy-test a 0.70 average accuracy for S8 and a 0.93 for S9. The You Only Look Once v5 (YOLOv5) algorithm was successfully applied to identify and count the target objects and display the result in the touch display. Based on the outcome for noise reduction, placing different kinds of boxes and boxes with the same size and dimension of the target objects will not be detected. YOLO is the best algorithm for object detection that recognizes only the trained objects. The results show the accuracy, which offers a high precision and high recall curve and decreases all lost when the dataset added more pictures, leading to higher accuracy.