Résumé: It is of great significance to realize high-precision detection of golden pomfret for intelligent management of fishery farming. Nevertheless, the highly variable size of the objectives and the degree of overlap between objectives make optimization of the algorithm challenge. To solve the problems mentioned above, we propose a golden pomfret detection algorithm that combines the improved transformer and the YOLOv5 framework to surpass not only the canonical transformer, but also the high-performance convolutional modules. The specific methods are designed as follows: (1) On the transformer frame, this paper designs a transformer with a progressively increasing number of cascaded tokens, that aims to improve detection accuracy by adaptively learning grid parameters based on the size of the golden pomfret in each image. To achieve a high-performance result, the large kernel convolution is included between the input image and feature space mapping. (2) Based on YOLOv5, we redesigned the prediction head to address different sizes of golden pomfret detection. Then, we replace the original prediction heads with deformable prediction heads to further improve network performance and training efficiency through fine-grained feature mapping of golden pomfret. In particular, the deformable convolution uses a novel generalized linear interpolation algorithm to reduce detection errors. (3) Considering the robustness of the network, we introduce the bags of useful strategies such as data augmentation and polynomial interpolation. Experimental results in the golden pomfret test set showed that the mAP is better than the original YOLOv5 network by 22.59%. Therefore, our algorithm can effectively detect golden pomfret in complex ocean scenes.

Langue: Anglais