In intensive aquaculture, the real-time detection and monitoring of common infectious disease is an important basis for scientific fish epidemic prevention strategies that can effectively reduce fish mortality and economic loss. However, low-quality underwater images and low-identification targets present great challenges to diseased fish detection. To overcome these challenges, this paper proposes a diseased fish detection model, using an improved YOLOV5 network for aquaculture (DFYOLO). The specific implementation methods are as follows: (1) the C3 structure is used instead of the CSPNet structure of the YOLOV5 model to facilitate the industrial deployment of the algorithm; (2) all the 3 × 3 convolutional kernels in the backbone network are replaced by a convolutional kernel group consisting of parallel 3 × 3, 1 × 3 and 3 × 1 convolutional kernels; and (3) the convolutional block attention module is added to the YOLOV5 algorithm. Experimental results in a fishing ground showed that the DFYOLO is better than that of the original YOLOV5 network, and the average precision was improved from 94.52% to 99.38% (when the intersection over union is 0.5), for an increase of 4.86%. Therefore, the DFYOLO network can effectively detect diseased fish and is applicable in intensive aquaculture.