Answers are grounded in this paper and cite their sources.
Accurate classification of crop cultivars is a crucial technical support for germplasm resource research and targeted breeding. Chrysanthemum has ornamental, beverage, and medicinal values, with over 30,000 cultivars worldwide. The large scale of germplasm resources and the subtle differences in key morphological traits such as flower shape and petal type make traditional manual identification difficult and inefficient. General deep learning models can achieve high recognition accuracy in chrysanthemum cultivar classification; however, their relatively large parameter size and computational overhead limit their efficient deployment on mobile or edge devices in field scenarios. To address these problems, we first constructed a large-scale fine-grained chrysanthemum dataset, Chry-FG-151, containing 43,325 RGB images of 151 cultivars across six color series (white, yellow, orange, pink, purple, red), captured under natural light using various types of mobile devices. To reduce the influence of complex field backgrounds during training, we developed a Background Substitution Module (BSM) that combines explicit foreground extraction, morphological refinement, and neutral solid-color background reconstruction to generate foreground-focused augmented samples. To strengthen spatial responses to flower-related regions, we integrated a Spatial Semantic Calibration Module (SSCM), implemented as a lightweight pointwise-convolution spatial gate without spatial downsampling, to recalibrate feature maps for fine-grained cultivar discrimination. To meet the computational constraints of mobile deployment, we built a lightweight classification network, CCNet, consisting of a stem, six inverted residual bottleneck layers, a last expansion layer, two DropBlock2D layers placed before and after SSCM, and a lightweight classifier. CCNet achieved a test accuracy of 99.31 ± 0.24% on Chry-FG-151 under the full training pipeline. Under the BSM-only public-dataset protocol, it achieved test accuracies of 92.82 ± 0.84% and 86.18 ± 1.76% on Public Dataset I and Oxford 102 Flowers, respectively. The model contains 0.16 M parameters and requires 0.10 G FLOPs. Grad-CAM visualization and foreground–background attention analysis showed that CCNet allocated a slightly larger proportion of activation responses to flower foreground regions than MobileNetV3-small. Across seven backbone models and two public datasets, the BSM-only setting yielded higher mean test accuracy than the no-augmentation baseline. The compiled model size was 312.13 KB, and the single-image inference latency on the tested mobile CPU was 143 ms. These results support the feasibility of on-device inference under the evaluated hardware setting. This study presents a lightweight, deployment-oriented approach for intelligent chrysanthemum cultivar recognition.
No takes yet. Share an insight, caveat, or question.
Feng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: