Abstract
Deep learning-based yuzu detection systems often suffer performance degradation when the training and inference datasets differ in collection sites, acquisition years, and imaging conditions. To address this issue, this study proposes a multi-loss self-reconstruction (MLSR) filter and an MLSR-based correction-detection framework.The proposed MLSR filter is an encoder-decoder architecture, where the input image is encoded into feature representations and then decoded back into the image space. It aims to reconstruct natural image characteristics while suppressing detection-irrelevant variations. To achieve this objective, we propose a staged optimization strategy with multi-loss constraints to train the filter, in which L1 and SSIM losses constrain the reconstruction to recover the basic image structure, the VGG perceptual loss constrains it to preserve detection-relevant features, and the PatchGAN adversarial loss prevents the reconstructed images from deviating excessively from natural image characteristics. In the correction-detection framework, the trained MLSR filter is applied to both training and inference datasets before yuzu detection. The proposed correction-detection framework was evaluated on UAV (Unmanned Aerial Vehicle) imagery collected under different orchards, years, and illumination conditions. The results showed consistent improvements for YOLOv5, YOLOv8, and YOLOv11, with maximum gains of 5.62% in F1-score, 7.72% in mAP50, and 6.95% in mAP50-90. We argued that MLSR filtering mitigated feature shifts between the training and inference datasets, allowing YOLO to capture more consistent patterns across them and thereby improving detection performance. However, this mitigation was reflected in YOLO backbones but did not introduce obvious visual changes. Therefore, we measured the similarity between the training and inference datasets based on YOLO backbone representations before and after MLSR filtering. The results showed that YOLO perceived higher feature similarity between the datasets at detection-relevant scales after MLSR filtering, consistent with the reported detection improvements.

This publication has 19 references indexed in Scilit: