Hybrid Deep Learning and Computer Vision System for Real-Time Detection and Color-Based Counting in Assembly Tasks

Abstract

In laboratory and low-cost industrial settings, automated detection and counting of small assembly components continue to be difficult tasks, especially when illumination conditions are unpredictable and annotated data is few.  In this paper, a hybrid computer vision system that combines a deep learning-based object detector with traditional colour analysis is presented for the real-time identification and color-based classification of assembly cubes.  The suggested method uses an HSV-based post-processing technique for colour recognition and a YOLOv8 model trained as a single-class detector for cube-shaped object localisation.  While retaining good detection and classification accuracy, this division greatly lowers dataset complexity and annotation effort. A GoPro Hero 8 camera running in webcam mode at 1080p resolution and 30 frames per second was used to experimentally evaluate the system. To increase resilience against changes in illumination, object orientation, and background, a new dataset of over 400 photos was employed for both training and validation. With mAP50 values ranging from 0.93 to 0.96, experimental results show good detection ability and dependable colour categorisation in a variety of test conditions. On consumer-grade hardware, real-time operation was attained at 15–25 frames per second. The results confirm that the proposed YOLO–HSV hybrid architecture offers an efficient, interpretable, and scalable solution for color-based object detection and counting. The system is particularly suitable for educational laboratories, rapid prototyping, and low-cost assembly monitoring applications, bridging the gap between deep learning research and practical deployment.