Enhanced Self-Attention-Based Rapid CNN for Detecting Dense Objects in Varying Illumination

This paper addresses the challenge of efficient detection of densely arranged unordered items under varying illumination. Specifically, a novel convolutional neural network-based method is proposed for item vector detection, recognition, and classification, termed Self-Attention and Concatenation-Based Detector (ACDet). In a benchmark pharmaceutical case study, rapid and accurate detection of pharmaceutical package contours is achieved, enabling the automatic and fast verification of both the quantity and types of pharmaceuticals during distribution. At the input stage, a combined image augmentation method is applied to improve the detection model’s ability to learn the appearance features of items from multiple angles. Based on YOLOv8 model, integrating computational module C2F with Attention (C2F-A), multidimensional self-attention reinforcement is applied to the outputs of multiple gradient streams. The designed Weighted Concatenation (WConcat) module self-learns to weight and concatenate multi-level feature maps, enhancing the model’s cognitive capability. Finally, simulation experiments are conducted to determine the optimal timing for utilizing each module. Simulation experiments compared the proposed ACDet with several state-of-the-art YOLO architecture models utilizing the benchmark Comprehensive Pharmaceutical Package Dataset (CPPD). ACDet achieved 81.0% mAP and 79.5% Smooth mAP on the CPPD dataset, outperforming other models by an average of 5.5% to 16.6%. On public datasets, the results were 52.2% and 51.0%, respectively. The impact of utilizing C2F-A at different stages on performance was also tested, concluding that the WConcat module does not necessitate spatial attention. Finally, in zero-shot testing, the verification success rate reached 99.91%. Our work shows that the proposed ACDet can overcome many challenges in complex object detection scenarios, enhancing robustness while maintaining a lightweight design. The proposed model can serve as a new benchmark.