Objective Dry matter and sugar content are two important indicators affecting the quality of kiwifruit. To achieve rapid and accurate detection of these indicators, a non-destructive detection method for key internal quality indicators of kiwifruit is proposed, integrating improved deep convolutional neural network with spectral technology.Methods A spectrometer was used to collect spectral data of kiwifruit, and the data were transformed into two types of two-dimensional images using Gramian Angular Field (GAF) transformation. An improved convolutional neural network model with multi-dilated convolutions was constructed to predict and analyze key quality indicators of kiwifruit. The model consists of two independent CNN modules connected in parallel to process the two types of two-dimensional images. Multi-dilated convolution strategies, clustering pruning methods, and channel attention mechanisms were incorporated to enhance the model's detection and analysis performance.Results Compared with other models, the proposed method reduced the average root mean square errors of dry matter and sugar content by 20.59% and 13.04%, respectively, increased the average determination coefficients by 6.45% and 4.34%, respectively, and improved the average relative prediction deviations by 6.99% and 12.78%, respectively.Conclusion The proposed method demonstrates good capability in detecting and analyzing key internal quality indicators of kiwifruit, and provides a valuable reference for non-destructive internal quality testing of kiwifruit.