Fruit Image Recognition with Enhanced CBAM Attention Mechanism Based on Resnet18

Abstract

In the field of computer vision today, the accurate recognition of fruit images is of great significance for practical applications such as intelligent agriculture and food quality inspection. Traditional image recognition models encounter a precision bottleneck when processing fruit images, primarily due to the diversity and complexity of fruit features, particularly when different fruit varieties exhibit similar appearances. This paper focuses on improving the performance of fruit image recognition models and proposes an innovative method based on the ResNet18 architecture, which integrates multiple pooling and spatial attention mechanisms. By constructing a dynamic pooling fusion module, the model can adaptively learn the weights of different pooling methods, thereby effectively alleviating feature conflicts among multiple pooling methods. At the same time, the introduction of a channel-spatial attention dynamic balance mechanism optimises the allocation of attention to features in different dimensions, enhancing the pertinence and effectiveness of feature extraction. Experimental results show that, compared to the original ResNet18 model, the average accuracy of the improved model on the fruit30 dataset has increased significantly from 73.933% to 75.108%. The highest accuracy on the fruit360 dataset has also increased, fully demonstrating the excellent effect of this method in enhancing fruit feature recognition ability and improving the model's generalisation performance. It lays a solid foundation for the broad application of fruit image recognition technology in practical scenarios.