通过图像增强与改进Faster-RCNN网络的重叠鱼群尾数检测

利用目标检测获取水下鱼类图像中的生物信息，对于实现水产养殖信息化、智能化有重要意义。受到成像设备与水下拍摄环境等因素的影响，重叠鱼群尾数检测仍为水下目标检测领域的难点之一。该研究以水下重叠鱼群图像为研究对象，提出了一种基于图像增强与改进Faster-RCNN网络的重叠鱼群尾数检测模型。在图像预处理部分，该研究利用MSRCR算法结合自适应中值滤波算法进行水下图像增强；在Faster-RCNN网络的改进部分，该研究采用ResNeXt101网络作为模型主干网络、增加带有CBAM（Convolution Block Attention Module）注意力机制的Bi-PANet（Bilinear-Path Aggregation Network）路径聚合网络、使用PAM（Partitioning Around Medoids）聚类算法优化网络初始预测框的尺度和数量、以Soft-NMS（Soft Non-Maximum Suppression）算法替代NMS（Non-Maximum Suppression）算法。通过以上措施提高模型对于重叠鱼群尾数的检测精度。通过消融试验可得，改进后的模型对水下重叠鱼群图像的平均检测精度和平均召回率分别为76.8%和85.4%，两项指标较Faster-RCNN模型分别提高了8.4个百分点和13.2个百分点。通过对多种模型的实际试验结果进行对比可知，改进后的模型的平均准确率相较于YOLOv3-spp、SSD300和YOLOv5x6分别高出32.9个百分点、12.3个百分点和6.7个百分点。改进后的模型对重叠数量为2～5尾的鱼群进行数量检测时，成功率分别为80.4%、75.6%、65.1%和55.6%，明显高于其他目标检测算法，可为重叠鱼群尾数检测提供参考。

Detecting overlapping fish population using image enhancement and improved Faster-RCNN networks

Populational number of fish is of great significance in the intelligent aquaculture. An accurate and rapid target detection has been always obtained from the underwater images. Much attention is still paid on the single target during underwater fish detection so far. However, the overlapping target is largely limited to the applicative scenes. In this work, an overlapping fish detection model was proposed using image enhancement and an improved Faster-RCNN network, named after CP-Faster-RCNN. First, 15 healthy grass carps were fed in a test tank for 15 days, where the images of fish movement were collected by the equipped underwater camera. The total 2 077 available images were filtered to serve as the dataset for the modelling. The training, test, and validation set were divided, according to the rate of 6:2:2, respectively. In the training set, all recognized fishes in each image were labelled by the rectangle bounding boxes via a labelImg tool. Next, three most commonly-used detection models (YOLOv3-spp, SSD300, and YOLOv5x6), as well as the new proposed model (CP-Faster-RCNN) were trained using the training and test set, and then were verified by the validation set. Among them, the CP-Faster-RCNN model was optimized for such the overlapping fishes in the images. Hence, some improvements were made using the original Faster-RCNN model. The details were as follows. 1) The MSRCR combined with the adaptive median filtering was adapted to improve the definition of underwater images; 2) The ResNeXt101 network was served as the backbone network to enhance the feature extraction capability of the model; 3) A specific bilinear path aggregation network (Bi-PANet) with the Convolution Block Attention Module (CBAM) attention mechanism was designed to fully use the multi-scale feature maps, which was effectively reduced the interference of background information. 4) The PAM clustering was chosen to optimize the scale and number of initial anchors of the network, which was used to speed up the convergence of the model; 5) The Non-Maximum Suppression was replaced by the Soft Non-Maximum Suppression, in order to reduce the missed detection for the overlapping objects. The results showed that the mean Average Precision (mAP) values of CP-Faster-RCNN were 32.9, 12.3, and 6.7 percentage higher than those of YOLOv3-spp, SSD300, and YOLOv5x6, respectively, indicating the best performance of detection. Besides, 300 extra images were randomly selected from the validation set to test the actual performance of the four models. Statistical analysis demonstrated that there was the different detection accuracy at the different number of overlapping fishes in clusters. Once the number was from 2 to 5, the mAPs of CP-Faster-RCNN decreased orderly, which were 80.4%, 75.6%, 65.1%, and 55.6%, respectively. More importantly, the mAPs of CP-Faster-RCNN was rather higher than that of other three models in the same way, indicating the better suitable performance in the scenes of overlapping fishes. Last, five ablation tests were carried out to investigate the specific effects of five improvements, where the mAP was the index of assessment. It was found that the mAP of the CP-Faster-RCNN was 76.8%, which was the totally 8.4 percentage increase than that of the original Faster-RCNN network. The backbone network was replaced for the 5.7% mAP increase, due to the more substantially complex model. Meantime, the pre-processing of dataset was also offered a 0.9% increase of mAP. The other three improvements for the Faster-RCNN posed the positive impacts on the model complexity, together contributing 1.8% increase for the mAP. All improvements were promoted the more accurate detection for the model. Therefore, the improved model can be expected to achieve the much more accurate fish detection, especially for the overlapping fish images in the complex underwater backgrounds.