Résumé:
This master’s thesis addresses the problem of automatic recognition of harmful insects, a crucial challenge for precision agriculture and food security. Faced with the complexity and slowness of manual identification, the main objective of this work is to develop and validate an intelligent, modular, and lightweight system capable of detecting and identifying insects in real-time for deployment on mobile devices. Our approach is based on a two-stage pipeline : object detection with the YOLOv12n model (a version of YOLOv12), followed by species classification with FastViT-SA12 (a variant from the FastViT family). The system was trained and evaluated on the complex IP102 dataset, which includes 102 classes and presents major challenges such as data imbalance and high visual variability among insects. The experiments have demonstrated the effectiveness of our approach : the detection module achieved an excellent mAP@50 score of 95.8% on the test set. The classification module, in turn, obtained a Top-1 accuracy of 73.77% and a Top-5 accuracy of 90.82%, results that are highly competitive and rival the state of the art. In conclusion, these results validate the relevance of our modular system as a pragmatic and high-performing solution. The balance achieved between accuracy and efficiency, notably a low mobile latency (130 ms), confirms its suitability for practical application in precision agriculture.
