Channel Pruning-Based YOLOv7 Deep Learning Algorithm for Identifying Trolley Codes

Abstract

The identification of trolley codes poses a challenge in engineering, as there are often situations where the accuracy requirements for their detection cannot be met. YOLOv7, being the state-of-the-art target detection method, demonstrates significant efficacy in addressing the challenge of trolley coding recognition. Due to the substantial dimensions of the model and the presence of numerous redundant parameters, the deployment of small terminals in practical applications is constrained. This paper presents a real-time approach for identifying trolley codes using a YOLOv7 deep learning algorithm that incorporates channel pruning. Initially, a YOLOv7 model is constructed, followed by the application of a channel pruning algorithm to streamline its complexity. Subsequently, the model undergoes fine-tuning to optimize its performance in terms of both speed and accuracy. The experimental findings demonstrated that the proposed model exhibited a reduction of 32.92% in the number of parameters compared to the pre-pruned model. Additionally, it was observed that the proposed model was 24.82 MB smaller in size. Despite these reductions, the mean average precision (mAP) of the proposed model was only 0.03% lower, reaching an impressive value of 99.24%. We conducted a comparative analysis of the proposed method against five deep learning algorithms, namely YOLOv5x, YOLOv4, YOLOv5m, YOLOv5s, and YOLOv5n, in order to assess its effectiveness. In contrast, the proposed method considers the speed of detection while simultaneously ensuring a high mean average precision (mAP) value in the detection of trolley codes. The obtained results provide confirmation that the suggested approach is viable for the real-time detection of trolley codes.