Pollen is a valuable proxy for reconstructing current and past vegetation. Automating identification and counting of pollen grains could greatly help palynologists, by increasing sample size, and thus their spatial and temporal resolutions, and standardizing methods and results. Several recent studies have already shown the potential of deep learning for automatic pollen recognition, especially for aeropalynology. Studies on pollen traps and fossil samples remain scarce, most probably because they contain many non-pollen particles and damaged pollen grains, increasing the difficulty of the task. Here, we test a new combination of last-generation deep-learning algorithms for automatic detection and classification of pollen from annual traps containing as many as 70 Mediterranean taxa, and many debris types. A total of 16 traps were collected each of three consecutive years in six locations in France. For each trap, one slide was mounted, and photographed partially with an automatic microscope. This operation produced 1,024 images of 204x204µm per slide, which contained a few pollen grains, that could be damaged, cut, or clumped, and many debris. We first trained YOLOv5 to detect the single category of pollen on 85% of 4,096 images (256 images per slide) containing 12,344 manually detected pollen grains. On the remaining 15% of the annotated images, the model left 0.7% pollen undetected, and falsely detected 12% of debris which are meant to be excluded by the subsequent classification. We then applied the model on the remaining 12,288 images and obtained 42,156 additional pollen grains. For the classification, we have trained so far ResNet50 on 85% of 8,000 manually identified pollen grains among 26 classes, made of a single or a few pollen taxa, and one extra class of debris. On the remaining 15% of the images, we obtained a class-mean accuracy of 0.73 with per-class accuracy ranging from 0.28 to 0.96. The best classification rates were obtained for taxa we had most images for (Pistacia sp., Quercus ilex, Lycopodium sp.). We are now improving the classification, e.g. by testing other algorithms, increasing the training dataset and/or through data augmentation.