In this paper, the BottleneckCSP-small module is introduced into the automatic detection YOLOv5s model, replacing the standard convolutional unit in the backbone network, and optimizing the computational efficiency by combining the depth-separable convolution with the partial convolution (PConv) of FasterNet. In the feature extraction stage, the SENet attention mechanism is integrated to strengthen the adaptive weighting ability of channel features and enhance the sensitivity of disease region localization. The results show that the average detection accuracy of the improved model for normal, damaged and scarred fruits is 98.3%, 98.6% and 98.5%, respectively. The mean average accuracy of the lightweight improvement only was 95.6%, and the mean average accuracy of the further improvement combined with the attention mechanism reached 96.5%. The improved YOLOv5s has excellent performance in the three metrics of precision, recall and average precision, such as faster speed and more stable convergence.
