Research On Fish Object Detection Method Based On Image Enhancement And Improved CIoU Loss

China is a big country in aquaculture,and the amount of aquaculture accounts for a large proportion of the total amount of aquaculture in the world.However,at present,the identification of fish species in China’s aquaculture mainly relies on manual work.Manual species identification is time-consuming and labor-intensive,and it is easy to make mistakes,which seriously limits the development of the aquaculture industry.At this time,a fast and accurate fish object detection system is particularly important.The traditional fish object detection method is to manually extract the shape,size,color,texture and other features of the fish,and then input the feature vector into the classifier for classification,but the traditional method uses fewer features and is difficult to analyze the large amount of data.The data set for feature extraction has great limitations.After decades of development,deep learning has made significant achievements in the image field,and its detection results are more accurate and robust.In this paper,a deep learning method is used for fish object detection,and a fish object detection method based on image enhancement and improved CIoU loss is proposed.The main research work of this paper is as follows:(1)Aiming at the problem of image blur caused by the influence of water quality,lighting and other factors in fish images captured in real breeding environments,a method of combining CycleGAN and IQA-CNN models to enhance images is proposed.First,the IQA-CNN model is used to evaluate the image quality of the three datasets in this paper.According to the evaluation results,the images are divided into two categories,one with an evaluation score higher than the threshold,and the other with a quality score lower than the threshold.The CycleGAN model is used to train those with scores lower than the threshold as the source domain,and those with scores higher than the threshold as the object domain for training to achieve the purpose of enhancing low-quality images.Finally,the YOLO v4 model is used to detect the data before and after the enhancement.The experimental results show that the detection results are greatly improved when the appropriate threshold is selected.The mAP reached 93.74% on the self-built data set in this paper,and the Fish4 Knowledge data set.99.41% mAP was achieved,and 91.68%mAP was achieved on the NCFM dataset.(2)Aiming at the problems of slow detection speed and low detection accuracy of existing fish object detection algorithms,a detection method with improved CIoU loss is proposed.A new loss term is constructed based on the CIoU loss function,so that the intersection of the predicted box and the ground-truth box is regressed in the same way as the aspect ratio of the ground-truth box during the regression process.When the real frame and the predicted frame do not intersect,the predicted frame will gradually approach the real frame due to the influence of other loss items,and the added loss items will start to work when the intersection starts,which will make the predicted frame close to the real frame in the process of x The incremental ratio of the axis and the y-axis is kept close to the width and height ratio of the real box,so that the width and height of the predicted box and the real box keep the same ratio of growth,which speeds up the convergence of the model.When the predicted box intersects with the ground-truth box but the aspect ratio of the intersected box and the ground-truth box is different,the predicted box is affected by the added loss term,which will first make the aspect ratio of the intersected box and the ground-truth box consistent,until it reaches the real box.The box has the same aspect ratio as the intersecting box.The improved CIoU loss is applied to the YOLO v4 model.Compared with the original model,the mAP is greatly improved.The mAP reaches 94.22% on the self-built dataset,99.52% on the Fish4 Knowledge dataset,and NCFM dataset.92.16%.The data enhanced with the above CycleGAN algorithm achieves 95.14% mAP on the self-built dataset,99.67% mAP on the Fish4 Knowledge dataset,and 92.44% mAP on the NCFM dataset.