**4. Conclusions**

The nutritional value of rice decreases with the fineness of the rice DOM, while the processing process causes unnecessary food waste and affects national food security.

The purpose of this study was to solve the problems of the high labor intensity of traditional manual detection of rice DOM with manual feature extraction and a low recognition rate of existing classification methods based on machine learning. This paper presents an IRBOA model capable of extracting multi-scale rice features to identify classified rice DOM to further guide the processing process of rice enterprises.

The classical CNN model was improved by fusing the Inception-v3 structure and the residual structure. IRBOA, a multi-scale information fusion model, was constructed and its identification accuracy was enhanced relative to other classical networks. In addition, we used the BOA to seek the hyperparameters that led to the optimal performance of the model and increased the correct classification rate of the model. The IRBOA model, which performed hyperparameter optimization by BOA, achieved a recognition rate of 96.90% for rice DOM, while the testing time for a single image was less than 20 ms. The accuracy of IRBOA improved by 7.41 and no less than 1.35 percentage points relative to traditional machine learning methods and classic CNN models, respectively. The model enhances the feature representation and has better classification performance and generalization ability.

This study has demonstrated the feasibility of the inspection method proposed, which can provide a certain guidance to the processing work of rice enterprises and provide a reliable and accurate technical means for the classification of rice DOM level. More importantly, real-time rice DOM level evaluation can be achieved in the actual production process. Subsequently, the model can be combined with specific sorting apparatus to sort rice that has reached a certain DOM level in the rice milling section. It avoids the rice being over-milled in the next milling stage, so as to reach the goal of moderate processing and grain saving.

However, there are still some shortcomings in the research of this paper, and we will improve our current work in the following two aspects in the future work: (1) The model is prone to error attributed to the acquisition of single-sided images due to the different bran degrees on two sides of different DOMs rice. In the future, we will adopt the method of double-sided image acquisition [38] to improve the recognition rate of the model. (2) The chalky region of rice will have an impact on the discrimination of DOM level. In future research, we will search for effective image processing means to reduce the influence of the chalky areas of rice. (3) The accuracy of the model proposed only reaches 96.90%, which not only takes a long training time but also requires a large number of training samples. In the future, we can try to use the lightweight model [39,40] with small samples to save training time, or use the transfer learning model [41,42] to improve the recognition accuracy while reducing training time and samples.

**Author Contributions:** Conceptualization, W.C. and W.L.; methodology, W.L.; software, W.L.; validation, W.L. and Y.W.; formal analysis, W.C.; investigation, W.L.; resources, W.C.; data curation, W.L. and Y.W.; writing—original draft preparation, W.L.; writing—review and editing, W.C. and W.L.; visualization, W.L.; supervision, W.C.; project administration, W.C.; funding acquisition, W.C. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** This study does not involve humans or animals.

**Informed Consent Statement:** This study does not involve humans.

**Data Availability Statement:** The data presented in this study are available on request from the corresponding author.

**Conflicts of Interest:** The authors declare that they have no conflict of interest.
