Abstract:Aiming at the high cost and low detection accuracy of heat exchanger plate assembly inspection,an improved YOLOv7 algorithm was proposed to detect the reverse installation of heat exchanger plate.On the original data set,the limited contrast adaptive histogram equalization was used to process the image to improve the clarity of the image to be detected; on the basis of the original YOLOv7 model,for the layers number of the heat exchanger plate is large,blurred and much noisy, a multi-level adaptive attention mechanism was proposed to add to the last layer of Backbone; in view of the large amount of calculation of the image recognition model and the serious loss of downsampling features,the improved MP-D module was used to optimize the original downsampling module; in the feature extraction part,the F-ReLU activation function was added to significantly improve the calculation speed and detection accuracy; for the PANet structure of the Neck part,the idea of cross-scale connection of BiFPN was integrated to further improve the efficiency and accuracy of fusion.Through experiments,it can be seen that compared with the original YOLOv7,the improved network model has mAP@0.5,recall rate R,and inspection speed increased by 0.6%,2%,and 16.9 frame per second,respectively.It has a good effect on the detection of heat exchanger plates.