Method for detecting Ophiocephalus argus cantor under intra-class occulusion based on cross-scale layered feature fusion

A method for detecting Ophiocephalus argus cantor under intra-class occulusion based on cross-scale layered feature fusion, comprising image collecting, image processing and a network model;

• wherein the image processing comprises dividing collected images into a training set and a prediction set according to 9:1, labeling all images with labelimg to obtain images of ground truth boxes containing all objects, and size-clustering all the ground truth boxes to obtain nine ground truth box sizes most suitable for Ophiocephalus argus cantor detection training, and adjusting image sizes to form input images, wherein the input images are suitable for the network model;
• inputting the input images into the network model, detecting the objects, extracting features of the object Ophiocephalus argus cantor from the input images by 1×1 convolution, and integrating the features, adjusting dimensions, inserting cross-scale layered feature fusion modules,
• which is characterized in that,
• firstly, performing a model training, inputting input images of the training set into the network model, wherein the input images of the training set are in the cross-scale layered feature fusion modules, dividing all features input into the modules into n layers composed of s feature mapping subsets in total, fusing features of each feature mapping subset with other feature mapping subsets, finally connecting to form complete information fusion, and after convolution operation, outputting training result containing object confidence, coordinate information and class information;
• then using a YOLOV4 loss function to adjust network parameters, and after 50 training iterations, obtaining parameters suitable for the network model to form a network model for detection;
• and
• then detecting the model, using input image of the prediction set as test images for detection, and inputting the test images into the network model with adjusted parameters, obtaining, by the network model, prediction result containing object classes, center point coordinates and width and height information of candidate boxes, and inputting the prediction result into a non-maximum suppression module, and screening, by the non-maximum suppression module, correct prediction boxes based on a ranking of locating accuracy scores of all candidate boxes.