Cross-Scale Feature Fusion for Occluded Snakehead Fish Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting Ophiocephalus argus cantor in underwater environments face challenges with intra-class occlusion due to the fish's slender and snake-like body, leading to missed detections and overlapping objects, especially in aquaculture settings where accurate detection is crucial.
Innovation Solution
A method utilizing cross-scale layered feature fusion, involving image processing, network model training, and non-maximum suppression to enhance detection accuracy by integrating features across multiple scales and adjusting for overlapping bodies, effectively extracting and distinguishing features of Ophiocephalus argus cantor under occlusion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional object detection methods are used to detect Ophiocephalus argus cantor, then detection speed can be maintained, but detection accuracy deteriorates due to intra-class occlusion and overlapping bodies
Solution Approach 1:
The patent divides the detection task into multiple scales by implementing feature fusion across different detection layers (P3, P4, P5). Each scale focuses on detecting fish of different sizes and occlusion levels, segmenting the complex detection problem into manageable scale-specific sub-tasks that improve overall accuracy without requiring a completely new model architecture
Solution Approach 2:
The patent merges feature maps from different detection layers through feature fusion modules, combining semantic information from deeper layers with spatial information from shallower layers. This merging of multi-scale features enables the model to detect occluded fish effectively while maintaining reasonable computational complexity through shared backbone networks
2Measurement precision
If multiple anchor points are set for image sub-blocks to detect multiple objects, then detection coverage improves, but false detection and missing measurement increase due to similar object sizes and coincident center points
Solution Approach 1:
The patent introduces scale as an additional dimension by implementing multi-scale detection. Instead of relying solely on multiple anchor points at a single scale, the model detects objects at different scales (P3, P4, P5 layers), adding a scale dimension to the detection space. This allows the model to distinguish overlapping fish by their relative sizes and positions across scales, reducing false detections and improving reliability
3Measurement precision
If standard feature extraction is used for slender-bodied fish, then computational efficiency is maintained, but feature discrimination deteriorates under diverse body postures and occlusion
Solution Approach 1:
The patent performs preliminary feature extraction at multiple scales during the forward propagation phase, preparing scale-specific feature maps (P3, P4, P5) before the detection stage. This preliminary organization of multi-scale features enables efficient subsequent processing and fusion, maintaining detection speed while improving feature discrimination for occluded fish through pre-computed scale-specific representations
Data Source
AI summary
Disclosed is a method for detecting Ophiocephalus argus cantor under intra-class occulusion based on cross-scale layered feature fusion, including image collecting, image processing and network model, where collected images are labeled, image sizes are adjusted to obtain input images, and the input images are input into an object detection network, integrated by convolution and inserted into cross-scale layered feature fusion modules, characterized by including dividing all features input into the cross-scale layered feature fusion modules into n layers, composed of s feature mapping subsets, and fusing features of each feature mapping subset with that of other feature mapping subsets, and connecting; carrying out convolution operation, outputting training result; adjusting network parameters by a loss function to obtain parameters for a network model; inputting final output candidate boxes into a non-maximum suppression module to screen correct prediction boxes, so that prediction result is obtained.