Improved Detection Method for SODL-YOLOv7 Intensive Juvenile Abalone

Achieving rapid and accurate detection of juvenile abalone is a prerequisite for estimating the number, density and size of juvenile abalone. Juvenile abalone are densely distributed in the breeding process, and the intra-class occlusion between each other is serious. Microorganisms in the water form inter-class occlusion for juvenile abalone, resulting in incomplete detection information. There is a lack of effective detection methods for juvenile abalone. To address the above problems, this paper proposed the SODL-YOLOv7 juvenile abalone detection method based on the establishment of the JAD (Juvenile abalone detection) dataset. First, the SODL backbone network for dense small target detection is proposed to improve the attention to small targets by incorporating null convolution kernels and pooling kernels with different sampling rates in the spatial null convolution and pooling layers; then, the ACBAM (Adaptive convolutional block attention module) is established to apply the adaptive pooling layer of channel space attention module, so that the network can pay more attention to the young abalone occlusion region and further improve the detection effect. Finally, the method of used in this paper was tested on the JAD dataset, with the results that the AP (average precision) reached 99.4%, an increase of 4.1% compared with the benchmark method YOLOv7, an increase of 9.2% compared with the instant-teaching method, and an increase of 2.2% compared with the TOOD method, therefore verifying the effectiveness of the method of this paper.