YOLO-ARM: An enhanced YOLOv7 framework with adaptive attention receptive module for high-precision robotic vision object detection

Abstract

This study addresses the difficulties of low detection precision, poor real-time performance, and poor model generalization in robotic vision systems under adverse circumstances through the proposition of an improved object recognition scheme based on a better convolutional neural network (CNN). To address these ends, YOLOv7-improved architecture is proposed, referred to as YOLO-ARM, which employs two new modules: the Adaptive Attention Receptive Module (ARM) and the Convolutional Block Attention Module (CBAM). ARM enhances feature extraction by adjusting the dynamic receptive field and multi-scale feature fusion, whereas CBAM improves feature maps by using channel and spatial attention procedures to improve the attention of the model towards critical features. The contributions of this paper involve the combination of ARM and CBAM in YOLOv7 to enhance the capacity of the model for handling scale changes, occlusions, and clutters. ARM module leverages group convolutions, squeeze-and-excitation blocks, and depth-wise convolutions for strengthening feature discrimination, while CBAM leverages channel and spatial attention in order to boost respective features. The proposed YOLO-ARM model outperforms other models on the MS COCO dataset, with an F1-score of 98.60 %, precision of 97.997 %, and accuracy of 99.727 %.