Optimization of Robot Vision Object Detection and Recognition Algorithm Based on Convolutional Neural Network

Abstract

Accurate object detection and recognition remain fundamental challenges in robot vision systems operating in complex environments. To improve detection accuracy, robustness, and computational efficiency, this study proposes a multi-stage collaborative optimization framework based on convolutional neural networks. A lightweight backbone architecture combining depthwise separable convolution and channel attention mechanisms is first designed to reduce computational complexity while preserving semantic representation capability. An adaptive feature pyramid network is then developed to enhance multi-scale feature fusion and improve small-target recognition performance. Furthermore, a GIoU-based localization optimization strategy and a joint knowledge-distillation–pruning framework are introduced to improve localization accuracy and model efficiency simultaneously. Experimental evaluations on public and self-constructed datasets demonstrate superior performance in terms of detection accuracy, real-time processing capability, and parameter reduction. The proposed framework provides an efficient solution for intelligent robotic perception and offers potential applications in machine vision, electromagnetic imaging interpretation, and autonomous sensing systems.