Conquering camouflage: a collaborative feature refinement detector for near-color backgrounds and small-size challenges.

Field detection of flat jujubes faces three coupled challenges: extreme color similarity between fruit and leaves, small fruit size, and the requirement of lightweight deployment on edge devices. Existing methods either address one of these problems in isolation or simply stack existing techniques without a principled design for the triple conflict. This paper proposes Collaborative Feature Refinement Detector (CFRDet), which synergistically integrates four modules to resolve the above conflicts. Mish activation preserves weak gradients, preventing small target features from vanishing in lightweight networks. The parameter-free Area-Attention Similarity-Aware Module (ASAM) suppresses near-color background interference without increasing parameters. A dedicated small target detection layer recovers spatial details lost during downsampling. Mobile Inverted Residual Bottleneck (MIRB) detection heads process the high-resolution feature map with nearly linear complexity, offsetting the computational cost of the small target layer. Together, they form a collaborative chain of "feature preservation, background suppression, detail recovery, low-cost computation". Experiments on our self-built flat jujube dataset showed that CFRDet improved mAP and F1 by 1.7 and 3 percentage points respectively over a strong baseline, reduced FLOPs to 5.6 G, and compressed the parameters and model size to 2.3 MB and 5.1 MB. Comprehensive ablation studies verify the necessity of each module and their synergistic gains. In comparisons with mainstream models, it surpassed small-sized models in accuracy while retaining lower complexity than nano-sized networks, and its generalization was further confirmed on a sweet persimmon dataset. The model's dual-attribute of precision and lightness underpins intelligent automation in field flat jujube harvesting.